Return to the source
Store verbatim evidence with project scope and provenance so a later answer can cite the record behind it.
Skill
Memory
Pick up the evidence, not just the summary.
Keep scoped source notes, find them in a later session and trace derived facts back to what was actually recorded. Start with a small explicit evidence CLI; add automatic capture only when you need it.
Project: fictional-parser Evidence: "Parser check remains unverified." Source: fictional-note Later retrieval: exact source text + content reference scoped to this project Current truth: recheck before relying on it
This fictional example shows the evidence trail. The native quickstart lets you test a real local save and fresh-process search with temporary fictional data.
Keep changes inspectable
Separate evidence from derived facts and navigation. Supersede a stale fact while retaining the evidence trail.
Start without a model
The explicit native evidence CLI saves and searches local notes without a host, background daemon or embedding model.
The film
The words just traveled.
In 64 seconds, a vendor promise turns into a myth by spring, and a houseplant named Fern pays for it. Then Memory keeps the first words with their project, call and date, hands them back weeks later, and links a newer fact to its source without throwing the old one away. Captions are on the player; a transcript is below.
Read the transcript
Monday. The vendor says: the sensor drifts after thirty days. By Thursday it’s: may drift. By spring: totally reliable! Nobody lied. The words just traveled.
On day thirty-one, Fern finds out. Rewind.
Memory keeps the first words in a jar: which project, which call, which day. Weeks later, ask what the vendor said, and you get the exact words back, with where they came from.
When a fact changes, the new one points to its source. The old one stays on the shelf.
On screen: Pick up the evidence, not just the summary.
Memory. Find the original words, later.
Fern, the vendor and the call are a fictional example. The film shows how the skill behaves; it does not report results.
See the work take shape.
This fictional example shows the evidence trail. The native quickstart lets you test a real local save and fresh-process search with temporary fictional data.
Keep one useful note
File a fictional parser-check note in a named project and room. Record where it came from rather than saving a conclusion without context.
Find it in a later process
Search the same native log for parser. Lexical/BM25 retrieval returns matching evidence and its reference without asking a model to reconstruct the past.
Review what changed
Read the source before treating an old claim as current. Derived facts cite evidence; later facts supersede earlier ones without erasing the original record.
Fictional example
Fictional worked example
The brief A fictional task illustrating Memory.
Before
A remembered conclusion is reused as current fact with no source.
After applying the method
Scoped evidence and a provenance-linked superseding fact preserve what was known and when.
Durable retrieval stays separate from unverified claims of current truth.
A clear job.
A useful stopping point.
Remember, retrieve, revise or inspect durable project evidence.
Use it when
- A later session needs the exact project note behind a decision.
- You need evidence, derived facts and navigation to stay distinguishable.
- You want scoped local records you can inspect and preserve.
Choose another route when
- You only need to pass the next task between agents: Project Handoff is a smaller starting point.
- You need physical deletion simply by invalidating a fact.
- You expect installing a method folder to activate automatic memory or migrate existing stores.
Remember
Evidence deserves durable retention.
Verbatim evidence scoped to its project and source.
Retrieve
A current task needs relevant past evidence.
Scoped source-linked results with currency qualified.
Maintain
A claim changed or retained records need review.
Explicit supersession, export or authorized forgetting.
Read the canonical method
Memory
Memory is a consumer of the separate amplifier-data storage capability. Keep that modular boundary: multiple capabilities may share the storage substrate, and this skill must not redefine or silently replace its AI-shaped contract.
Remember: retain task-relevant verbatim evidence with explicit wing/room scope and source/session/revision metadata when available. Do not capture incidental secrets or personal material merely because it appears in context. Automatic capture is an optional host integration, not a requirement of this method.
Retrieve: constrain searches to the relevant project and room. Return source passages and distinguish evidence from distilled facts and navigation indexes. A result recorded earlier is not an observation of current reality. Inspect the source before relying on a remembered claim whose currency matters.
Maintain: read references/reflection-rubric.md for evidence-to-fact distillation. Derived facts require provenance. Supersede stale facts without deleting the evidence trail. For an explicitly selected Project Memory SQLite schema-1 store, read references/legacy-archive.md and use its stdlib exporter/new-log archival import; other legacy formats remain unsupported. Export or migrate only through a documented adapter and explicit destination; never discover and convert an existing store as an installation side effect. The native log is append-only: invalidating a fact does not erase its bytes. For an authorized deletion request, inspect the actual retention/export procedure and identify separate backups; do not claim invalidation is physical deletion.
Read references/storage-boundary.md before runtime setup. Run scripts/runtime_doctor.py to inspect native availability without opening any user store. The portable method needs no database to install. Persistent execution requires a supported amplifier-data distribution. The narrow memory-evidence CLI uses the retained NativeMemoryStore directly without a model, daemon or host; full automated capture and briefing require the optional host adapter. Doctor must report an unavailable substrate honestly; it cannot make an unknown storage dependency disappear. Historical retrieval figures describe their original recorded protocol only, not this consolidation.
Limits and prerequisites
- Portable method installation is independent; native persistence depends on the separate amplifier-data substrate and supported adapters.
Your first useful run.
Choose the portable guidance or the native quickstart. Save one fictional note first, then verify you can find the same words in a fresh process.
The download contains the complete skill folder. Extract it into the skills directory supported by your agent, keeping its references and license together. Or use the source installer below. Python 3.11+ is required for that installer.
Start from the public source
git clone https://github.com/michaeljabbour/amplifier-skill-memory.git
cd amplifier-skill-memorypython3 scripts/capability.py install --dest ~/.agents/skillsNative tools and optional execution
python3 -m pip install 'git+https://github.com/michaeljabbour/amplifier-tool-data@40450647367db2a2b5058bea1d81a6d248c24c94'
python3 -m pip install --no-deps ./modules/tool-memory
python3 skills/memory/scripts/runtime_doctor.pyThese commands are for the optional native surface. Follow its installation guide and prerequisites; portable method installation is separate.
Keep the complete skill folder together, including its references and license. The host applies the method in your context; the skill does not start a background service.
- The portable skill is guidance. It installs without a database or Memory runtime.
- For a working native save/search check, follow docs/NATIVE-QUICKSTART.md. Python 3.11+, Rust and a native compiler are needed for source Data installs.
- The reviewed public Data library stays independent. Existing stores are never discovered or migrated by installation.
Try this with your agent
Use Memory to review this project note. Preserve the exact source, its project scope and the date. Separate observed evidence from any derived claim. Show what needs rechecking before we rely on it today.Use after installing in a host that loads skill folders. Supply your own task files and review the result.
Keep the guide nearby
Read the guide · License · Changes · Release metadata · CONSOLIDATION · LEGACY STORES · NATIVE QUICKSTART
Know what you can rely on.
The public native path is checked with temporary evidence, two scoped consumers, regeneration and durable restart. Full embedding, host and live-task acceptance remain separate; historical retrieval results keep their original protocol limits.
Read the checks and their scope
- Functional acceptance — passed
573 native module tests passed, 4 skips; actual native two-consumer and no-host evidence checks also passed. The reviewed code-only Data distribution is public; narrow native evidence acceptance is checked separately.
- Native host acceptance — not-run
No live model, real desktop transaction, invitation, deployment or production data migration was performed by this functional baseline.
- Storage ownership and runtime — passed
Public amplifier-data 0.2.0 Rust kernel resolves without credentials; original private source/history and all existing stores remain untouched.
Source and conformance checks are separate from demonstrated benefit. Comparative effectiveness and outside-user outcomes are not established by this site.
Fictional examples explain the capability. Tests check the stated behavior. Outside-user outcomes need their own evidence.