Blog

Memory, context and working with AI.

Analysis, comparisons and experiments on memory, context, continuity and decisions over time.

Explore the eighteen observatory articles and proposed methods.

A decision linked to its rationale, a rejected alternative and a limited exception.

Memory & Context

Thin graphite plates engraved with fine lines float in a calm space; some fade out of the frame and one amber thread links only a few of them.

Why does ChatGPT forget what I told it?

ChatGPT can remember useful details but does not keep the full state of a long project. Understanding chat, memory and working context is the key.

On the left, a cloud of small grey blocks converges on a single result; on the right, a line of blocks, some solid and some translucent, linked by the same amber thread.

AI memory vs RAG: they are not the same thing

RAG retrieves useful knowledge for a query. AI memory maintains evolving context over time. They may share retrieval technology, but they solve different problems.

A close view of four matte slabs stacked from ivory down to dark graphite, with a thin amber line filling a clean seam on each slab while the layers underneath stay visible.

The real test of AI memory starts when it is wrong

Useful memory is not memory that never fails. It is memory you can correct, update and forget without losing the history that still matters.

A wide grey and black band comes in from the left, passes behind two translucent plates, then splits into two branches that spread apart toward the right, each followed by an amber thread.

Can AI memory trap you in your past decisions?

Remembering past choices can help AI, but it can also create conservative bias. Useful memory should inform without deciding for you.

Two stacks of graphite slabs: one inert, the other threaded by an amber line that lights only a few of them.

AI memory vs context: storing information is only half the problem

An AI memory that stores facts does not understand a story. See what separates storage, retrieval and context, and when a simple memory is all you need.

A frosted glass frame holding a few glowing points, while a long trail of copper points stretches behind it and fades into the distance.

Why a great prompt cannot include what you forgot to mention

A careful prompt can only hold what you remember. See why the history of a project goes missing from your requests, and how to bring it back when it matters.

A thin copper line runs from a grey stone block engraved with fine lines to several glass markers set in a row; an identical block beside it, with no line, stays isolated.

Why preserve the reasoning behind a decision?

A decision written down without its reasons tends to get applied at the wrong moment. A simple template for keeping the why, and knowing when to revisit.

An amber lantern lights a single pebble among nearly identical grey pebbles; graphite threads connect the lantern to each one, and only one amber thread is pulled taut.

Why found does not mean true

When an AI retrieves a memory, that proves nothing by itself. Four questions about source, date and status to ask before you rely on what it reports.

Two overlapping glass planes: the older one pale and set back, the newer one sharper with an amber bead, linked by a copper thread.

What should happen when a decision changes?

Erase, pile up or replace: what each choice does to a project’s memory, and why a replaced decision must not turn into a false truth later on.

A row of glass candles: the first burn with a clear amber flame, the next ones are out, and a single one has been swapped for a new candle linked to the old one by a copper thread.

How to tell that a memory has become obsolete

Replaced, contradicted, expired, out of context or moot: five ways a memory goes stale, the signals to watch for, and a ten-minute review you can try.

Five different glass objects (sphere, cube, prism, ring, cone) lined up on a light surface, each tied to the base by a line of a different thickness and material: graphite, copper, dotted.

The difference between a memory, a fact, a decision, a hypothesis and an instruction

One event can turn into five very different notes. How to tell a memory, a fact, a decision, a hypothesis and an instruction apart, and avoid the mistakes.

A cloud of fine grey particles passes through a graphite sieve; only three small pebbles settle, one with an amber glow.

Why a good memory must also know when not to write

Keeping everything feels safe, but a memory full of noise and outdated facts helps less. Why not writing is a feature in its own right, not a gap.

Working with AI

Two abstract structures of different shapes, one rounded and one angular, joined by the same bundle of amber threads.

Switch AI Assistants Without Losing the Why Behind Decisions

Make decisions, sources, rights and outcomes portable across AI assistants without sharing the entire history or creating vendor lock-in.

Two glass crates: in the first, pages stacked with no connection; in the second, the same pages joined by graphite threads, with a copper thread leaving the crate and passing through a doorway.

Memory Should Be Portable Before It Is Intelligent

Evaluate AI memory portability before advanced features: exports, sources, rights, structure, history and recovery by another AI.

Several copper lines converge at a fork; past points are lit, but the decision ring at the center stays empty and unlit.

The Post-AI Lens: Test an Idea Beyond the Code

Seven questions to test an idea after code becomes cheaper: distribution, workflow, trust, data, learning and organizational leverage.

A thin copper line runs from a grey stone block engraved with fine lines to several glass markers set in a row; an identical block beside it, with no line, stays isolated.

The Decision-Outcome Journal: The Missing Loop in AI Teams

Connect every important decision to its hypothesis, expected outcome and observed result to create an organizational memory that learns.

A small closed glass chest holding a copper thread, set beside a large open graphite structure. A single fine, chosen thread passes from the chest to the structure through an opening.

One Memory Across Multiple Projects Without Mixing Everything

Organize AI memory across companies and projects with explicit boundaries, private annexes, shared references and authority.

A cloud of fine grey particles passes through a graphite sieve; only three small pebbles settle, one with an amber glow.

What AI Should Forget: Build a Memory Policy

Decide what AI memory should never retain, expire, correct, restrict or delete based on purpose, risk and authority.

Dark stone milestones on a pale surface, crossed by one continuous amber thread, through zones of light and texture that suggest abstract seasons.

How to keep AI project context for months

A practical method for preserving decisions, current state, hypotheses and observed outcomes when you work with AI on a project for weeks or months.

Five identical copper wire loops, each on a small stone block, gradually merge into one calm, readable line.

How to stop repeating the same information to AI

Stop rebuilding the same AI context by separating stable information, current project state and the temporary details of the task in front of you.

A heap of raw shards on the left, then a clean cube, ring, prism and sphere linked by a short amber thread.

What should AI remember from a meeting, project or decision?

AI should not necessarily memorize an entire meeting. Durable memory should focus on decisions, reasoning, commitments, open questions and outcomes that will matter later.

Several copper lines converge at a fork; past points are lit, but the decision ring at the center stays empty and unlit.

How AI can retrieve your past trade-offs without deciding for you

AI can reread your past choices to help you think clearly, without deciding for you. Here is how to keep the judgment on your side of the table.

Comparisons

Two abstract storage systems side by side: a stack of soft layers and a block of project compartments, crossed by one fine amber thread.

ChatGPT memory vs Claude memory: what is different?

ChatGPT and Claude both support memory, but they organize continuity differently. A dated October 2026 comparison of chat history, project memory and portability.

A sober triptych: a cloud of points drawing together, a block of stacked sources and a wavy line of continuity, all crossed by one amber thread.

RAG, vector database or AI memory: which one do you need?

A decision guide for choosing document retrieval, vector search or long-term AI memory based on the actual problem your system needs to solve.

Three abstract architectures side by side, with no ranking: a stack of modular blocks, a graph of nodes on an axis, a layered stack of discs, joined at the base by a discreet amber thread.

Mem0, Zep and Nawame: different approaches to AI memory

Mem0, Zep and Nawame all address persistent AI context from different starting points. A practical comparison of infrastructure, context engineering and user-owned memory.

A long block of pale stone with stacked grey layers widens toward the right, while a thin amber line runs straight through it and ends on a small dot against an ivory background.

When not to use Nawame

Nawame is not useful everywhere. Native memory, README, Notion, short projects or systems of record: when a simpler tool is enough.

A small sealed glass jar holding a short copper thread, and to the right a longer thread passing freely through several glass containers of different shapes.

Why not simply use ChatGPT or Claude memory?

ChatGPT or Claude built-in memory is often enough. Here is when it is, when a provider-independent history makes a difference, and where Nawame falls short.

A large glass vat packed with tight grey pieces, where a single amber thread is barely visible. Beside it, a small glass cup holds just three sharp pieces joined by a clear amber thread.

Giant prompt or targeted context?

Should you paste your whole history into the prompt? Cost, noise, contradictions: when a giant prompt is fine, when targeted context wins, and how to choose.

Three glass containers of different shapes on a light tray: one sealed, one resting on a stone base, one open and joined by a long copper thread to a fourth that floats above.

Local memory or independent memory?

Built-in memory, files on your computer, a home-made knowledge base or an independent memory: four approaches compared, with their trade-offs and their limits.

A perspective grid of matte graphite cards, with a fine copper line crossing several of them to one card glowing amber.

Notion, Obsidian or Nawame: documenting is not the same as remembering

Notion and Obsidian document very well. Remembering decisions over time is a different job. Where each tool shines, and how to combine them.

Privacy & Portability

Two abstract structures of different shapes, one rounded and one angular, joined by the same bundle of amber threads.

How to move project context from one AI to another

Moving from ChatGPT, Claude or another assistant without starting over takes more than exporting chats: transfer state, decisions, reasons, constraints and provenance.

Grey pebbles stay on one side of a filtering glass partition while a thin bundle of amber threads passes through into an empty glass chamber.

How to share AI context without sharing personal data

Useful AI context does not require exposing your entire personal memory. Separate scope, identity, sensitive data and working context before retrieval.

A stack of grey plates dissolves into fine dust on its right side, while an amber line keeps running steadily from left to right across the whole picture.

What if Nawame disappeared tomorrow?

Portable memory is more than an Export button. What should remain understandable if the service itself disappears?

An amber thread tangled with small black strands struggles against a thick dark plate, in front of thin translucent panes, and never reaches the large charcoal block behind it.

How AI memory can become a prompt injection channel

Retrieved memory must remain data, not silently become privileged instructions for an AI system.

An opaque graphite block stays in place while a bundle of copper threads with small beads flows out of a glass case across the frame.

Your data belongs to you: what should a real AI memory export contain?

An Export button is not enough. Content, useful history, replacements, provenance, versioned formats: what an export needs to stay useful once you leave.

Two glass crates: in the first, pages stacked with no connection; in the second, the same pages joined by graphite threads, with a copper thread leaving the crate and passing through a doorway.

Data portability vs portability of meaning

Downloading your conversations does not let you switch AI without losing the thread. What separates a raw export from one that keeps meaning, and how to test it.

Two glass cubes, one clear and one frosted, each threaded by a copper line, joined by a fine balance line.

Sovereignty and AI memory: why privacy can require trade-offs

More privacy sometimes means fewer features. What Nawame’s sovereign mode changes, what it costs, and how to choose depending on how you work.

An open glass box with grey sheets coming out of it; each sheet passes under a small magnifying glass, and only a few get an amber thread that leads into a stone filing cabinet.

Importing memory without importing hallucinations

A memory exported from another assistant is full of summaries and guesses, not guaranteed facts. Here is how to sort it before you bring it in.

Teams & Organizations

Three flat discs float one above the other: a cream white one, a charcoal one with fine concentric rings, and a last one made of grey layers. An amber thread links them through a golden dot on each.

AI Memory Without Sources Is Just a Persistent Opinion

Reliable AI memory preserves source, date, confidence, rights and observed outcome : not just a convincing summary.

A row of glass candles: the first burn with a clear amber flame, the next ones are out, and a single one has been swapped for a new candle linked to the old one by a copper thread.

Why Organizational Memory Needs to Know Information Freshness

Accurate information can become wrong. Learn how to date, review and expire items in an organizational AI memory.

Five different glass objects (sphere, cube, prism, ring, cone) lined up on a light surface, each tied to the base by a line of a different thickness and material: graphite, copper, dotted.

Conversation, Decision and Outcome: Three Objects You Should Not Confuse

A conversation explores, a decision commits and an outcome verifies. Separate these objects to build actionable organizational AI memory.

Grey pebbles stay on one side of a filtering glass partition while a thin bundle of amber threads passes through into an empty glass chamber.

Forgetting as a Feature of AI Memory

Reliable memory must correct, expire, unlink and delete information without destroying useful history or legitimate obligations.

An amber lantern lights a single pebble among nearly identical grey pebbles; graphite threads connect the lantern to each one, and only one amber thread is pulled taut.

How to Store an Inference Without Turning It Into a Fact

Prevent hypotheses from becoming truth through repeated summaries by preserving source, epistemic status, contradiction and review.

Abstract illustration of connected memory and decisions

Shared Memory, Private Memory and Authority: Drawing the Boundaries

Reliable memory architecture separates private notes, shared team knowledge and decisions that carry authority.

Several empty compartments of glass and dark stone, partitioned apart, joined by a few context points and amber threads.

How to build company memory that AI can actually use

Useful company memory is not one giant document repository. AI needs facts, decisions, rules, history, provenance and access boundaries to be kept distinct.

A twisted copper wire, a braided graphite cord and a translucent filament enter one glass sphere without blending.

Team AI memory: how to keep sources and contexts separate

Shared AI memory becomes risky when decisions, personal opinions and different projects are mixed together. Scope, provenance and authority need explicit boundaries.

A compact block of four stacked plates, from ivory to charcoal, surrounded by large translucent sheets and an amber thread that circles it without touching it.

What AI memory should never copy from your system of record

CRM, ERP, Git or banking data should not become duplicated truth inside AI memory. Memory should preserve the context those systems do not explain.

Three flat discs float one above the other: a cream white one, a charcoal one with fine concentric rings, and a last one made of grey layers. An amber thread links them through a golden dot on each.

Source, validation and authority are three different things

Knowing where information came from does not mean it is true or authoritative. Organizational memory has to keep those statuses separate.

A small closed glass chest holding a copper thread, set beside a large open graphite structure. A single fine, chosen thread passes from the chest to the structure through an opening.

Why personal memory must not automatically become company memory

Pouring everyone’s AI conversations into a shared memory looks efficient. Here is what gets lost, and how to share what matters without exposing the rest.

Two threads, one copper and one graphite, each pull a ribbon toward the same glass block. The block carries two readable lines on top of each other, and a third pale thread links them without merging them.

What happens when two people edit the same memory?

Two versions of one fact, a contradicted decision, a silent overwrite: how a team keeps a shared memory trustworthy, and what no tool can settle for you.

Lab

Two separate sources remain connected to their versions, with a visible correction and disagreement.

Reliable memory keeps sources, dates and disagreements

A source establishes origin, not necessarily truth. How sources, effective dates and corrections shape trustworthy assistant memory.

Shared context reaches several tools through distinct access boundaries.

Portable memory needs clear sharing boundaries

Changing assistants should not erase context. Why portable memory must preserve private work, access rights and shared sources.

An experimental loop turns observation, a test and verification into the next question.

A negative result can improve AI memory

Hypotheses, synthetic corpora, costs and measurement errors: how the Nawame Lab turns experiments into architectural decisions.

Different wordings lead to a shared note whose relevance still needs checking.

Semantic search: what embeddings can retrieve

Embeddings connect different ways of expressing an idea. Nawame’s research explains why similarity must remain separate from evidence and authority.

A question leads through evidence to another source without losing the wider case.

Why a handover needs more than one search

A first result may reveal what to search for next. Nawame’s experiments examined multi-step retrieval, evidence coverage and the limits of synthetic handovers.

A timeline separates recording, decision and effective dates.

Past decisions need their own time frame

Finding an old version is only part of historical retrieval. Nawame tested explicit validity periods to help assistants use the right decision.

A preserved source document linked to derived assertion cards.

Should AI memory be rewritten for machines?

Nawame compared human documents with structured assertions. The experiment separated compression gains from evidence quality and handover weaknesses.

The same value appears in two texts with different decision status.

Rewriting a decision can change its status

An assistant can preserve a number while turning a proposal into a rule. Nawame’s experiment explains why decision status must stay connected to its source.

Finding, citing and comparing evidence are separate stages.

Why an assistant can cite evidence and miss its consequence

Nawame tested guidance placement and question-specific alerts. Citation, comparison and justified abstention proved to be different capabilities.

A tool form distinguishes omitted fields, applied filters and partial results.

An optional parameter can hide the right memory

A Nawame test failure showed how tool contracts shape AI answers. Information can be present while an unintended filter makes it unreachable.

Unchanged fragments retained between two document versions while edited fragments are renewed.

Embeddings: recompute only what changes

How reusing unchanged document fragments reduces embedding computation, and what our Lab result actually measures.

A queue separates document-chunk processing from a short query’s waiting time.

Choosing embedding infrastructure through comparable measurements

Shared or dedicated CPUs: findings from a Nawame benchmark, queueing effects and the limits of performance comparisons.

A selected passage stays connected to a qualification farther along in its source document.

Finding a decision inside a long document

Targeted passages can improve retrieval from long documents, but selecting them can lose qualifications or expose hidden instructions.

Memory cards retain their provenance and a role distinct from an approved instruction.

A memory is data, not an instruction

Why distinguishing memories, intended instructions and prompt injection is difficult, and what Nawame's classification experiments taught us.

A single source supports relationships and reading paths whose added complexity needs evaluation.

AI memory: when SQL is enough, and when to test a graph

What experiments with relationships, graph storage and vector indexes reveal about choosing an AI memory architecture.

Computation, continuity, recovery and permissions surrounding a memory service.

Resilience and full cost: what a memory service needs to measure

Computation, failover, recovery and encryption: understanding the full cost and trust boundaries of a memory service.

Decision memory linked to a business object and reused across several interfaces.

Memory for business software, even without a chatbot

How object-linked memory preserves decisions and reasons, with permissions, history and clear limits to the current Embedded evidence.

Selected events lead to a few explanation requests and reusable memory.

Capturing explanations without overloading teams

When to request a human explanation, how to qualify it, and why restrained prompts matter to useful business memory.

Three ways of accessing context compared for the same business question.

With and without memory: what an Embedded evaluation measures

Results from two model families on synthetic data: object-based retrieval, scoring errors and the limits of the conclusions.

A graduated glass ruler lying along a row of grey marbles: a few marbles carry a small amber spark, and clean empty slots break up the row.

Why Reliable Agents Need Traces, Evals and Human Recourse

A reliable agent exposes its journey, is evaluated on real cases and allows people to stop, correct and appeal its actions.

A compact block of four stacked plates, from ivory to charcoal, surrounded by large translucent sheets and an amber thread that circles it without touching it.

OpenTelemetry and AI Memory: Correlate Without Recording Everything

Connect technical traces, decisions and agent outcomes without turning observability into an intrusive organizational memory.

A glass hourglass whose sand partly climbs back up an amber thread; beside it, a pile of small grey loops, the redone tasks, is lower than the taller pile next to it.

Measure Tokens per Outcome, Not Tokens per Conversation

Connect tokens, human time, external tools and correction to a business outcome to optimize the true economics of an AI workflow.

Five flat pebble-like shapes stand in a row, from opaque charcoal to translucent white, close together without touching, while an amber thread weaves between them toward a small golden bead.

Agent Graceful Degradation: Continue, Pause or Ask

A practical method for deciding whether an AI agent should continue with limits, ask, queue work, hand off or stop.

A straight stone path leads to a clear target; beside it, a second path visibly detours around an obstacle. An amber thread marks the honest path, and dull marbles, the failures, sit on a visible shelf.

A One-Off Evaluation Does Not Protect an AI Agent From Drift

Build an evaluation schedule that detects changes in models, data, tools, use and populations after production release.

Two identical structures made of clear and dark panels face each other across a smooth floor, joined only by a few fine threads, one of them amber.

Declared Trust Is Not Evidence

A framework for checking AI trust claims through observable controls, tests, outcomes, limitations and appeal.

Five flat pebble-like shapes stand in a row, from opaque charcoal to translucent white, close together without touching, while an amber thread weaves between them toward a small golden bead.

A good memory must know when to retrieve nothing

Finding something is not always better. Reliable memory has to know when no memory is relevant enough to use.

Two identical structures made of clear and dark panels face each other across a smooth floor, joined only by a few fine threads, one of them amber.

Can an AI judge whether another AI remembered correctly?

AI judges can evaluate many answers quickly, but memory quality cannot be reduced to one automatic score.

Nearly identical glass beads connected by fine graphite lines; only one amber line runs unbroken back to a graphite anchor point on the left.

When semantic search finds a precedent that never existed

Semantic search can surface a memory that resembles your question without actually being about it. Why that is risky, and how to account for it.

A thin glass funnel above a large heap of grey fragments. Only a few fragments pass through, joined by an amber thread, while a mass of graphite threads tangles at the bottom.

Limits of retrieval at scale

Why more documents can mean worse answers: recall, precision, near-duplicates, freshness, contradictions, and what you can do about each of them yourself.

A copper line crosses an empty space and ends at an empty glass ring, with fine graphite tick marks below.

Why we deliberately test topics that do not exist

Asking a memory a question with no answer checks that it can say "nothing" instead of making something up. Why the test matters and what it does not prove.

A fine glass sieve above a tray: a few items fall through and are joined by an amber thread, while most stay caught in the sieve with no thread.

How to test non-writing

A reliable AI memory must know what not to keep. A six-step, repeatable test with eight made-up events, counting false positives and false negatives.

A graduated glass ruler lying along a row of grey marbles: a few marbles carry a small amber spark, and clean empty slots break up the row.

How we measure memory errors

False recall, omission, an old fact taken for current: seven kinds of AI memory errors, how to catch each one, and why a single score misleads.

A straight stone path leads to a clear target; beside it, a second path visibly detours around an obstacle. An amber thread marks the honest path, and dull marbles, the failures, sit on a visible shelf.

How to test an AI memory without cheating on the benchmark

Known test sets, cherry-picked cases, metrics changed afterwards: how people fool themselves with AI memory benchmarks, and a method to avoid it.

A glass hourglass whose sand partly climbs back up an amber thread; beside it, a pile of small grey loops, the redone tasks, is lower than the taller pile next to it.

How to measure whether a memory really avoids rework

Is an AI memory useful or just reassuring? A repetition log kept over a few weeks can tell you, with no special tooling and no invented numbers.

A long chain of glass links; a link with a faint amber tint near the start is repeated by several threads in the links that follow, while a pale correction thread replaces it without erasing the old one.

The errors a long-term memory can carry forward

A mistake written down once can be read a hundred times. Five ways it spreads through a long memory, and the habits that make it visible.

A thin glass spiral whose turns pass through the same point; an amber thread leaves it at a tangent at one spot, against a calm, neutral background.

Can we detect decision cycles without psychologizing the user?

A question reopened with nothing new is something you can state with dates, not with a diagnosis. Neutral wording rules, examples and limits.

A glass disc with a fine hairline flaw revealed by amber light, surrounded by measurement rings and tick marks.

Why we publish Nawame’s limitations too

False precedents, noise, dilution and time: here is what Nawame does not yet do well, what we measure so far, and why we choose to say it openly.