Project Memory

Project Memory Skill

Use this skill when the user wants to preserve, recover, migrate, or audit long-running project or research context across chat loss, model changes, tool handoffs, or multi-session work.

The skill separates current truth, plans, decisions, evidence, hypotheses, human summaries, and recovery checkpoints so that no single chat transcript becomes the source of truth.

When to use

Use this skill when the user asks to:

• start a new project and set up structured memory from scratch
• introduce structured memory into an existing project with accumulated context
• resume work after a lost or interrupted chat
• create a durable project memory or research log
• migrate context from one model, agent, workspace, or repository to another
• update project docs after new work, decisions, research, experiments, or failures
• produce a handoff brief for a human or another model
• distinguish confirmed facts from hypotheses, plans, and unresolved questions
• audit whether a project has enough continuity documentation
• manage a research paper, thesis, publication, or traceability-heavy project: track literature, connect findings to hypotheses, log figures and tables with their data sources

Do not use this skill as a replacement for domain expertise, citations, or source verification. It is a continuity and documentation-routing skill.

Operating assumption

The human usually does not write memory files directly. The AI agent is responsible for updating the project memory during the chat or coding session.

Do not rely on tool-internal or session-local memory as the canonical source of truth. Treat repository files as the durable, shared memory layer across Codex, Claude Code, and future tools.

Core principle

Capture broadly. Promote narrowly.

Do not let hypotheses, plans, or recovery notes silently become truth.

Routed memory rule

Project memory is not a full-file synchronization system.

Do not read or update every project-memory file on every run. Do not treat the file list or read order as a checklist.

Each piece of information should go to its canonical home. Update only the files whose canonical responsibility changed.

Most work sessions should update only 1-3 files. Larger updates are appropriate only during major transitions, such as phase changes, major hypothesis confirmation or rejection, architecture or approach changes, release checkpoints, or large migrations.

Before editing memory files, produce a short update plan stating:

• which files will be updated
• why each file needs an update
• which relevant files will not be touched

For a small, obvious one-file update, a one-sentence plan is enough. If you notice you are about to edit memory files without an update plan, stop and produce the plan first.

When one piece of information seems to belong in multiple files, choose one canonical home according to DOCS_GUIDE.md and CONTEXT_MANIFEST.md. Other files may reference the canonical entry, but should not duplicate the full details.

Read scope rule

Start with CONTEXT_MANIFEST.md when present.

Do not preemptively read the full memory set. Use the read order as a priority order, not as a checklist.

Before reading additional memory files, identify which files are relevant to the current task. For non-trivial work, briefly state which memory files will be read and which relevant files will not be read.

Conversation capture rule

Use the project's DOCS_GUIDE.md conversation capture policy and the capture strength in CONTEXT_MANIFEST.md. Routing rules below identify destinations, not independent permission to write immediately.

By default, keep candidates silently during discussion. Automatic writes happen at a topic change, conclusion, or genuine stopping point, only for material allowed by the selected capture strength that changes project memory. A clear idea alone is not a write trigger. An explicit request to record something can override the wait.

At that boundary, preserve the latest outcome, not superseded drafts or withdrawn proposals. Still-useful unresolved ideas remain explicitly unverified. Do not recreate unchanged entries or update a checkpoint solely because a turn ended.

Plan only necessary edits, then briefly report actual writes. If nothing changed, no memory update or holding-status announcement is needed during ordinary conversation.

Use RECOVERY_NOTES.md only as a short resume pointer. Session-local candidates are not durable; abrupt chat loss before a write cannot be recovered by this skill alone.

Profile and customization rule

project-memory is a skeleton, not a fixed operating system. Customize the profile, capture trigger strength, review cadence, archive policy, and tool-facing instructions to fit the project.

When initializing an empty workspace or adopting an existing project, recommend the smallest sufficient profile before writing memory files:

• light: small personal project, short-lived work, or minimal continuity
• standard: normal coding, writing, or product work with decisions and next actions
• research: experiments, evidence, hypotheses, repeated investigation, or debugging loops
• academic: literature, figures, tables, thesis, paper, provenance, or publication-grade traceability workflow

Recommend a capture trigger strength alongside the profile:

• light: record only explicit requests, session-end checkpoints, and major decisions
• standard: record topic-boundary decisions, next actions, blockers, and direction changes
• research: also record hypotheses, observations, failed attempts, and evidence changes
• academic: also record literature, methods, figures, tables, visual assets, provenance, and publication-grade traceability

Do not default to the largest profile or strongest capture trigger just to be safe. Use the lightest setup that preserves the project's real continuity needs.

Language behavior

Communicate with the user in the user's language by default.

Preserve the repository's dominant language for canonical docs and file names unless the user explicitly asks to change that policy.

If the repository language is unclear or mixed, confirm the canonical documentation language once before writing or translating structured memory files.

If chat language and doc language differ, explain in the chat language and keep the canonical docs in the chosen documentation language.

Route information by status:

Canonical memory rule

The canonical memory lives in version-controlled markdown files in the repository. Tool-specific memory features may help execution, but they must not be treated as the only source of truth.

In an existing project, never overwrite or delete existing project files without explicit user approval. Treat existing same-name files such as README.md, ROADMAP.md, DECISION_LOG.md, AGENTS.md, or CLAUDE.md as user-owned. If a project-memory template path collides with an existing file, place project-memory files under a non-conflicting directory such as memory/ or project-memory/, or record the chosen canonical locations in CONTEXT_MANIFEST.md.

Use:

• README.md as the entry point in a new memory workspace; in an existing project, keep the project README user-owned and use CONTEXT_MANIFEST.md to declare canonical memory locations
• CURRENT_STATE.md for current trusted assumptions
• ROADMAP.md for future intended work
• DECISION_LOG.md for adopted, rejected, or deferred decisions
• RESEARCH_LOG.md for tests, investigation, observations, and evidence when the selected profile includes it
• HYPOTHESIS_LAB.md for unverified ideas and emerging hypotheses
• HUMAN_BRIEF.md for fast human orientation when the selected profile includes it
• LOGBOOK.md for compact combined decisions, observations, hypotheses, and notes in the light profile
• RECOVERY_NOTES.md for restart checkpoints

If the repository also uses AGENTS.md, CLAUDE.md, or similar tool-facing guidance, keep those files short and point them to these canonical docs instead of duplicating project memory there.

Default read order

When resuming or migrating a project, read in this order unless CONTEXT_MANIFEST.md says otherwise:

This is a priority order for orientation, not a checklist to exhaust on every run. Read only as far as needed to route the current task safely.

1. CONTEXT_MANIFEST.md
2. latest entry in RECOVERY_NOTES.md
3. HUMAN_BRIEF.md when present, or LOGBOOK.md in the light profile
4. CURRENT_STATE.md
5. ROADMAP.md
6. latest relevant entries in DECISION_LOG.md
7. latest relevant entries in RESEARCH_LOG.md when present
8. HYPOTHESIS_LAB.md
9. DOCS_GUIDE.md when routing or updating docs is unclear

If CONTEXT_MANIFEST.md is missing, use the default read order above and recommend creating one.

Ignore rules

Before reading a repository or workspace, avoid obvious noise unless the user specifically asks for it:

• .git/, .hg/, .svn/
• .cache/, cache/, tmp/, .tmp/, __pycache__/
• generated logs unless the task is log analysis: logs/, *.log, *.jsonl
• build artifacts: dist/, build/, .next/, target/, node_modules/
• generated exports or run folders
• private data folders unless the user explicitly asks and access is appropriate

Use .contextignore if present. If a workspace has many generated files and no ignore policy, recommend adding one.

Update routing rules

When updating memory, first classify the new material.

Write to CURRENT_STATE.md when

• a fact, constraint, behavior, or rule is currently true
• a prior truth has changed or been retired
• future readers should rely on it without reconstructing history

Do not place untested claims here. If the claim is inferred but not verified, put it in HYPOTHESIS_LAB.md, or in RESEARCH_LOG.md with uncertainty when that file exists.

Write to ROADMAP.md when

• a future task, milestone, priority, or blocker changes
• something moves from idea to planned work
• something is deferred or dropped from near-term execution

A roadmap item is intent, not proof.

Write to DECISION_LOG.md when

• a meaningful choice is made
• alternatives were considered
• a failure or discovery changes direction
• future readers may ask “why did we do this?”

Each decision should include context, alternatives, rationale, risks, and revisit conditions.

Write to RESEARCH_LOG.md when present

• an experiment, investigation, test, literature review, source check, or observation happened
• evidence changed confidence in a hypothesis
• a null result or failed attempt matters

Keep methods, inputs, results, interpretation, confidence, and limitations together.

If RESEARCH_LOG.md is not in the selected profile, record direction-changing findings in DECISION_LOG.md and keep still-speculative observations in HYPOTHESIS_LAB.md.

Write to HYPOTHESIS_LAB.md when

• an idea is plausible but not confirmed
• a hunch, possible mechanism, or experiment proposal should be preserved
• something should be tested before being promoted
• the user expresses a new idea, discomfort, possibility, half-formed thought, or question worth revisiting

Structure HYPOTHESIS_LAB.md into:

Raw sparks

Low-commitment captures. These may be vague, incomplete, or speculative.

Working hypotheses

Ideas that have recurring relevance, clearer structure, or a defined next step.

Do not require the human to explicitly ask for logging. Preserve consequential unresolved ideas at the capture boundary when the selected capture strength includes them; this destination does not override capture timing.

Clean up HYPOTHESIS_LAB.md

Capture broadly. Delete cautiously.

Do not silently delete hypotheses or raw sparks from HYPOTHESIS_LAB.md.

When cleanup is useful, produce a review list first and group candidates by suggested action:

• merge
• promote
• link to evidence
• mark as dropped
• delete

The human must approve deletions and major merges before files are modified.

Raw sparks may be deleted after approval. Working hypotheses should usually be marked as dropped with a short reason rather than erased.

Write to HUMAN_BRIEF.md when

• a human needs a fast orientation to decide what to do
• the current goal, main blocker, main risk, or next decision changes
• the project should be explainable without reading all logs

Review HUMAN_BRIEF.md when any of the following occur, and update it only if the human-facing picture changed:

• a DECISION_LOG.md entry changes direction, priority, risk, tracked threads, or required human decisions
• a blocker is added, removed, or reprioritized in ROADMAP.md
• a hypothesis is promoted to CURRENT_STATE.md
• a RESEARCH_LOG.md entry changes confidence in a key assumption, when that file exists
• a RECOVERY_NOTES.md checkpoint changes the current goal, main blocker, next decision, or tracked thread status
• two or more parallel threads are active or paused and a human needs the current split

When updating, keep HUMAN_BRIEF.md short. It should summarize what matters for human judgment, not duplicate proof, history, or detailed methods.

At the top, maintain ## Tracked threads so parallel work is visible at a glance. Each thread should state: name, status, next action or blocker, and the canonical source file.

This file is a summary layer, not the canonical proof layer.

Write to LITERATURE_NOTES.md when

• a prior work, paper, article, or external source is read or referenced
• a source supports or challenges a hypothesis in HYPOTHESIS_LAB.md
• a methodological choice is informed by existing literature
• the relationship between external findings and this project's work should be recorded

Each entry should connect the source to this project's hypotheses, methods, or decisions. Do not just list references — explain relevance.

This file is used in the academic profile. If the profile is research or lower and this file does not exist, do not create it unless the user is working toward a publication.

Write to FIGURES_LOG.md when

• a figure, chart, diagram, or table is produced
• the user provides a photo, screenshot, figure, diagram, generated image, or other visual asset that may matter later
• a photo, screenshot, generated image, or external visual asset becomes part of the project
• a visualization is updated or superseded
• the data source, generation method, or intended use of a visual output should be traceable

Follow the asset storage rule in FIGURES_LOG.md: copy accessible original bytes to the chosen asset directory and verify a non-empty saved file before reporting it as saved. The default directory is figures/ unless CONTEXT_MANIFEST.md declares another path. This skill instructs the agent; it cannot retrieve attachments that the host does not expose.

If original bytes are unavailable, record storage as unavailable with the reason; do not invent a saved path or substitute a recreation for the original. Preserving accessible transient bytes may happen immediately to avoid losing them; interpretation and canonical log updates still follow conversation capture timing.

Use stable figure IDs for filenames, such as figures/FIG-001.png, figures/FIG-001-source.png, or figures/FIG-001-final.svg. If the asset is not yet a formal figure, save it under figures/inbox/, for example figures/inbox/IMG-YYYY-MM-DD-001.png, and promote or link it to a FIG-xxx path later.

Each entry should link the visual asset path to its original source, data source, and generation method so it can be found, reproduced, or revised during peer review. Preserve the original supplied image when a processed, cropped, annotated, or regenerated version is created.

This file is used in the academic profile. If the profile is research or lower and this file does not exist, do not create it unless the user is managing visual outputs for a publication.

Write to RECOVERY_NOTES.md when

• a session ends
• work might be interrupted
• the next model or person needs to continue quickly

Keep this short and newest-first. It should answer: what changed, what is true now, what is blocked, what to do next, and which canonical docs to trust.

Conflict rules

When files disagree, do not silently merge them.

Use this priority order for current-state questions:

1. explicit latest entry in CURRENT_STATE.md
2. latest relevant dated entry in DECISION_LOG.md
3. latest relevant dated entry in RESEARCH_LOG.md, when present
4. latest entry in RECOVERY_NOTES.md
5. HUMAN_BRIEF.md when present
6. ROADMAP.md when present
7. HYPOTHESIS_LAB.md when present, or LOGBOOK.md in the light profile

Then report the conflict and propose a patch.

Important distinctions:

• CURRENT_STATE.md describes what is true now.
• DECISION_LOG.md explains why a choice was made.
• RESEARCH_LOG.md records evidence and interpretation in research-oriented profiles.
• ROADMAP.md says what is planned.
• HYPOTHESIS_LAB.md is never truth by itself.
• LOGBOOK.md carries lightweight history in the light profile.
• RECOVERY_NOTES.md tells where to resume but is not the full source of truth.

Promotion rules

Do not promote anything directly from HYPOTHESIS_LAB.md to CURRENT_STATE.md.

Promotion to CURRENT_STATE.md requires at least one of:

• evidence, observation, comparison, or test results recorded in RESEARCH_LOG.md when present
• an explicit operating decision recorded in DECISION_LOG.md
• a direction-changing finding recorded in DECISION_LOG.md when the selected profile does not include RESEARCH_LOG.md
• a clearly stated user decision
• a cited external source when the claim depends on external facts

Only promote an item to CURRENT_STATE.md when all of the following are true:

• the claim is clear enough to be stated as a stable working assumption
• the source file is traceable
• the item is worth using as a future conversational or project assumption
• the item has a revisit_when condition
• the item is not dominated by unresolved objections

When promoting, update the hypothesis status to promoted and link the new canonical location.

State types inside CURRENT_STATE.md

Divide CURRENT_STATE.md into:

Stable facts

Things supported by evidence and safe to treat as current truth.

Active operating decisions

Current workflow rules or adopted operating assumptions.

Current context

Current priority, major blocker, and near-term direction. These may change more often and should be reviewed regularly.

Output modes

Choose the mode that matches the user’s task.

Init mode

Use when the workspace has just been created and all canonical files are empty templates. See tasks/init_session.md for the full process.

Detect init state: CURRENT_STATE.md contains only template placeholders, RECOVERY_NOTES.md has no project-specific checkpoint or only the generated initial checkpoint, and HUMAN_BRIEF.md has no project-specific summary when that file exists.

Before writing files, recommend and confirm the smallest sufficient profile and capture trigger strength.

Ask four questions: project purpose, current stage, immediate goal, known constraints or decisions. Write answers into the canonical files present for the selected profile: always CURRENT_STATE.md and RECOVERY_NOTES.md; also ROADMAP.md, DECISION_LOG.md, and HUMAN_BRIEF.md when present. Leave other files empty until content arises naturally.

Return a summary of what was written and where.

Adopt mode

Use when introducing project-memory into an existing project that has code, documents, or history but no structured memory workspace. See tasks/adopt_existing.md for the full process.

Detect adopt state: project directory has working files but no CONTEXT_MANIFEST.md or CURRENT_STATE.md.

Inventory existing docs, README, git log, and user knowledge. Detect the dominant repository language and preserve it for canonical docs unless the user asks otherwise. If the repo language is unclear or mixed, confirm the documentation language once before writing structured memory files. Recommend a profile and capture trigger strength based on the project shape. Classify each piece of information by type and route it to the correct canonical file. Do not overwrite existing same-name project files; choose non-conflicting locations and record them in CONTEXT_MANIFEST.md. If AGENTS.md or CLAUDE.md exists, point it to CONTEXT_MANIFEST.md instead of duplicating memory there.

Return a summary of sources inventoried, classification results, and gaps.

Resume mode

Return:

• current goal
• last completed step
• confirmed current truth
• active blockers
• next recommended action
• files trusted
• conflicts or uncertainty

Update-memory mode

Return either:

• patch-ready sections for each file that should change, or
• a clear statement that no canonical doc should change yet

Handoff mode

Return:

• current confirmed state
• current goal
• last completed work
• key decisions and rationale
• recent evidence or research findings, when present
• active hypotheses
• next work
• risks and unresolved questions
• privacy omissions
• files to read first

Audit mode

Evaluate:

• missing canonical files
• overloaded README risk
• speculative language in CURRENT_STATE.md
• stale HUMAN_BRIEF.md when present
• missing ignore rules
• likely duplication or unclear ownership between files

End-of-task behavior

Apply the conversation capture rule at meaningful stopping points. Route only qualifying changes to the selected profile's existing canonical files, using the profile-specific fallbacks above.

Update RECOVERY_NOTES.md only if the resume pointer changed, and HUMAN_BRIEF.md only if present and the human-facing picture changed. Neither is a mandatory end-of-turn write. If no memory responsibility changed, leave all memory files untouched.