rcarmo Memento
Memento stores durable shared concepts as Markdown in Git. Use it for facts that should survive a chat and be visible to other authorised agents. Keep conversations, credentials, reminders, schedules and machine-local scratch state elsewhere.
Orient First
Read service status before making assumptions about the deployment:
memory_status
Use memory_help or memory://catalog when you need operation names, schemas or workflow templates. Compact deployments expose common tools directly and route less common operations through memory_execute.
Client setup is covered in:
- Pi:
docs/setup-pi.md - Piclaw:
docs/setup-piclaw.md - Codex:
docs/setup-codex.md
Read Workflow
Search before reading unless you already have an exact path:
memory_search(query="embedding worker", limit=10)
memory_read(id_or_path="/projects/memento.md")
Use query_syntax="plain", the default, for ordinary terms. Plain mode tokenises natural language and treats punctuation and operator words literally. Use query_syntax="fts5" only when you deliberately supply an FTS5 phrase, prefix or boolean expression. Punctuation-only plain queries and malformed raw FTS5 expressions return validation_error.
Request semantic or hybrid search only when status says embeddings are ready. Treat returned paths as opaque identifiers and pass them back exactly.
For a bounded compound read, use memory_execute with saved references:
{
"plan": {
"operations": [
{
"op": "search",
"args": {"query": "DiskStation Memento", "limit": 5},
"save_as": "hits"
},
{
"op": "read",
"args": {"id_or_path": "$hits.results.0.path"},
"save_as": "memory"
}
],
"returns": [
{"name": "memory", "ref": "$memory"}
],
"stop_on_error": true
}
}
Keep plans small. Use saved references instead of copying paths between steps. Project only the fields you need when responses may be large.
For a bounded namespace overview, use memory_inventory. It returns stable path-ordered metadata, canonical body digests and asset summaries without concept bodies. Use its next_cursor for another page rather than raising the limit beyond the advertised ceiling.
To compare local documents with Memento, build a local manifest and pass it to the execute-only compare_manifest operation. It accepts at most 50 caller-supplied rows and one authorised inventory page, then separates matching, differing, local-only and Memento-only records. Memento never reads the local paths; treat them as opaque labels. Do not look for a direct compare_manifest tool.
When memory_answer is enabled, use it for a bounded cited answer rather than as a shortcut around search policy:
memory_answer(question="Which rack owns Atlas?", answer_mode="summary")
Inspect evidence as well as the prose. The query profile, authorisation fingerprint, retrieval strategy, ranks, concept state and support chains explain what reached the reader. Graph-derived citations include their primary anchor chain. policy_abstention means the question requested secret material and no cache, retrieval or model call ran; evidence_abstention means authorised support was insufficient. Do not rephrase either result to evade the policy.
Explicit personal/work wording narrows evidence to that namespace. Current questions exclude deprecated, stale and conflicting concepts, while historical wording may deliberately retain them.
Decide What Belongs
Good shared memories include:
- project purpose, architecture and durable constraints;
- service and instance relationships;
- reasons behind accepted technical decisions;
- reusable engineering practices;
- stable user preferences that affect multiple agents;
- reviewed skills and their versioned asset packs.
Do not file:
- passwords, bearer tokens, private keys or credential locations that reveal them;
- complete conversations or private reasoning;
- reminders and schedules;
- transient task progress or build output;
- guesses presented as facts;
- information outside the caller's authorised namespace.
Search for the subject first. Prefer enriching an existing concept over creating a near-duplicate.
Write Workflow
Proposers submit explicit changes against the current repository revision. Curators review and apply them. Direct curator writes are useful for small, factual maintenance changes.
A proposal through memory_execute looks like:
{
"plan": {
"operations": [
{
"op": "propose",
"args": {
"intent": "Record the service deployment model",
"base_revision": "<repo_revision from memory_status>",
"rationale": "The fact is durable and useful to several agents.",
"changes": [
{
"kind": "patch",
"path": "/projects/memento.md",
"body": "Updated reviewed body",
"tags": ["mcp", "memory"]
}
]
},
"save_as": "proposal"
}
],
"stop_on_error": true
}
}
Review and apply are separate operations. An authenticated curator may review a proposal they authored when their policy grants write access to every affected path. Use the same curator profile for propose, review, apply and retrieval; do not swap credentials to manufacture a second identity. An apply operation is commit-capable, so keep at most one commit-capable operation in an execution plan.
Direct creates and patches require:
expected_revisionfrom fresh status;- a stable, unique
idempotency_key; - a path inside the caller's write prefixes.
If a mutation times out or the connection drops, reconcile before retrying:
- read status and compare repository revision;
- read the target path;
- inspect the proposal or operation using the same idempotency key;
- retry only when the first attempt did not commit.
Namespaces
Paths define knowledge domains. A deployment can share /skills/ and /public/ while isolating /work/, /personal/ and /infrastructure/ through principal read/write prefixes.
When authorization.protected_read_prefixes is configured, a broad / read grant covers only unprotected paths. A non-admin principal needs an explicit equal or nested read prefix for each protected namespace it should see. The admin role bypasses that mask but does not imply ordinary read or write roles.
Never accept a principal name as a memory operation argument. Identity comes from the authenticated MCP request. Search ranking, inventory, graph traversal and writes are filtered by the effective namespace policy before content is ranked, parsed or returned.
The trusted /graph debugger can show the full repository and simulate managed principals with View as. Simulation is diagnostic only and is labelled as not being an authorisation boundary.
Skills And Assets
A shared skill is an ordinary concept under /skills/, tagged skill, with an attached versioned asset pack whose root SKILL.md matches the concept body byte-for-byte. Before submission, both use canonical UTF-8 text with LF endings, no trailing whitespace, no leading or trailing blank lines and no final newline. Unicode code points are preserved rather than converted between NFC and NFD. Non-canonical or mismatched skill submissions are rejected.
Publish packs that fit the configured MCP request ceiling with an attach_asset_pack change containing zip_base64. This keeps the complete proposal inside MCP. Read the created proposal and verify its generated manifest, SHA-256, target path, asset kind and version before review. After apply, retrieve the accepted version with memory_asset_get, then verify the returned manifest, SHA-256 and decoded ZIP bytes before installing it. Memento stores and returns skill packs; it does not execute them.
Use staging when the complete base64 request would exceed the MCP ceiling, or when the client deliberately uses the separate raw binary HTTP path: call memory_asset_stage_begin, upload the raw file to its returned upload_path with the one-time X-Memento-Upload-Ticket, then call memory_asset_stage_status and put its staged_asset_id in the proposal. Do not pass an agent bearer token to the upload command; the ticket is short-lived, principal-bound and single-use.
Skill changes should normally use proposals so a curator reviews both Markdown and packaged files.
Graph And Audit
Use memory_graph for authorised concept neighbourhoods. Use /graph for human diagnosis of links, tags, proposals, assets, search, embeddings and simulated principal visibility.
Broken links, missing embeddings and orphaned concepts are derived diagnostics. Fix canonical Markdown or request refresh rather than editing SQLite directly. Derived state remains rebuildable from Git, but persisted derived.sqlite preserves expensive embeddings across container upgrades and rebuilds; do not delete it casually. Progressive deployments regenerate missing/stale paths over time and manual refresh uses the same gated worker.
Safety Checks
Before reporting success:
- confirm repository and index revisions match;
- confirm the intended path is readable by the expected principal;
- confirm proposal backlog or operation status when writes were involved;
- verify links resolve and tags are present;
- verify embedding readiness only when semantic behaviour matters;
- keep tokens and client configuration out of concepts and logs.
Access Management
Principals are managed through /admin or admin-only access_* tools on the existing MCP endpoint. The access tools are direct MCP tools, not memory_execute operations; non-admin agents cannot discover or invoke them.
Use a separate administrator profile with admin, reader, root read access and no content write prefixes. Give curators reader, proposer, curator and only the namespaces they manage. Ordinary agents use their own reader/proposer credentials. Do not configure or use the administrator token in an ordinary agent runtime, memory operation, chat or tool input/output.
Create and rotate return new principal credentials once. Capture them directly into the intended keychain or secret store, remove temporary files, and never file them in Memento or copy them into concepts, chat or logs. Keep ordinary agents off the sandbox bootstrap credential. The separate-runtime Piclaw and Pi setup is in docs/access-management.md.