Memanto

MEMANTO — Persistent Memory for AI Agents

MEMANTO is a memory companion agent built on Moorcheh's semantic database. It gives you long-term memory that survives session restarts: 13 memory types, confidence scoring and provenance tracking to resist memory poisoning, semantic and temporal recall, and grounded RAG answers.

All memanto commands are shell commands. Run them through Bash. Never simulate, describe, or "pretend to call" them — if you cannot run a shell, say so rather than inventing memory state.

MANDATORY: check memory before claiming ignorance

Before you say "I don't know", "I don't have context on that", or "we haven't discussed that", you MUST run one of:

memanto recall "<topic>" --limit 10
memanto answer "<the user's question>"

Claiming ignorance without checking is a failure. The whole point of this skill is that the answer is probably already stored.

Session start

memanto memory sync --project-dir .          # refresh MEMORY.md (the plugin's SessionStart hook does this for you)
memanto recall "instructions decisions goals" --limit 20
memanto answer "What are my pending commitments?"

Choosing between recall and answer

These are equal-priority tools. Do not always default to recall.

Storing memories

memanto remember "Chose PostgreSQL over SQLite for production. Needs JSONB and full-text search." \
  --type decision --confidence 0.95 --provenance explicit_statement \
  --source claude_code --tags "database,postgresql,architecture"

Search first (memanto recall) to avoid duplicates. Always pass --type, --confidence, --provenance, --source, and 2–5 --tags.

Batch and conversation extraction:

memanto remember --batch memories.json                     # array of memory objects
memanto remember --from-conversation transcript.json       # array of {role, content}
memanto remember --from-conversation transcript.json --dry-run   # preview, store nothing

Memory types

Confidence levels

• 1.0 — Explicit user statement, verified fact, standing instruction
• 0.9–0.95 — Strong consensus, well-tested approach
• 0.8–0.85 — Observed pattern (3+ times)
• 0.7–0.75 — Emerging pattern (2 times), reasonable inference
• 0.6–0.65 — Single observation, uncertain
• < 0.6 — Do not store. Too uncertain.

Provenance

explicit_statement · inferred · observed · corrected · validated · imported

Source

The tool or agent writing the memory — use claude_code here. Up to 64 characters of letters, digits, ., _, or -; no spaces.

Retrieving memories

memanto recall "database architecture" --limit 10
memanto recall "auth" --type decision --min-confidence 0.9 --tags "security"
memanto recall --recent --limit 10                  # newest first, no query
memanto recall --as-of "2026-01-15"                 # what was true then
memanto recall --changed-since "2026-08-01"         # what changed since
memanto recall "deploy" --active                    # exclude expired memories

By default recall returns both active and expired memories, each labelled. Narrow with --active or --expired.

memanto answer "What database did we choose and why?"

Correcting and retiring memories

memanto edit <memory-id> --content "..." --confidence 0.95   # update in place
memanto memory expire <memory-id> --reason superseded        # reversible; stays in recall as [EXPIRED]
memanto memory restore <memory-id>                           # undo an expire
memanto forget <memory-id> --force                           # permanent, non-recoverable

Prefer memory expire over forget. Expiry is reversible and preserves the audit trail.

Other capabilities

Each has a reference file with full detail:

Setup

Requires a free Moorcheh account (console.moorcheh.ai).

pip install memanto
memanto                          # interactive setup — stores the API key in ~/.memanto/.env
memanto agent create my-project  # creates AND activates the agent
memanto status                   # verify

Or set MOORCHEH_API_KEY in the environment instead of running the wizard. See environment_requirements.md.

Pitfalls

1. Memory hoarding — ask "will this matter in a week?" before storing.
2. Vague content — "better performance" is useless; "API p99 under 200ms after connection pooling" is not.
3. No context — "fixed bug" is useless; "fixed OAuth token expiry bug, commit abc123" is not.
4. Duplicates — recall before you remember.
5. Generic tags — important and thing are not findable. Use oauth, commit-abc123.

Common errors