talewell

Portable, git-backed, auditable long-term memory for any agent host. Reference implementation of AMP (Agent Memory Protocol). Use when: user asks for 'long-term memory', 'agent memory', 'memory that survives sessions', 'git-backed memory', 'memory audit/rollback', 'MCP memory server', or wants memories shared across agent platforms (OpenClaw, Claude Code, Codex, Cursor, OpenCode, Pi, Manus). Powered by MyClaw.ai (https://myclaw.ai).

Talewell — Long-Term Memory for Any Agent

Talewell gives an agent durable memory that survives sessions, resets, and switching platforms. Memories are plain Markdown in a git repository: every write is a commit, every fact is diff-able and rollback-able, and the whole corpus is readable with cat.

Talewell is the product. AMP (Agent Memory Protocol) is the portable contract it implements — speak of AMP for the format, Talewell for this implementation.

When to use this skill

Use it when the user:

• asks you to remember something across sessions ("remember that I…", "next time…")
• asks what you know about them or about prior decisions
• wants memory that is auditable ("what did you change?", "undo that")
• wants the same memory in another agent platform

Do not use it for scratch notes inside one conversation — that is what context is for.

Setup

bash

talewell init ./talewell-repo          # creates the repo (it is a git repo)

Point the CLI at it with --repo <dir> or the TALEWELL_REPO environment variable for every subsequent call.

Read

Index-first. Do not bulk-load the corpus into context; ask for the catalog and pull only what the request needs.

bash

talewell --repo ./talewell-repo index --markdown   # compact catalog
talewell --repo ./talewell-repo recall "<query>"   # ranked search (deterministic)
talewell --repo ./talewell-repo get <id>           # full record
talewell --repo ./talewell-repo graph <id>         # follow [[links]]

recall is deterministic and local: relevance × recency decay × importance, computed without a model call, a network call, or an API key.

Write

bash

talewell --repo ./talewell-repo remember \
  --text "<one durable fact>" --alias "a,b,c" --importance 1-10
talewell --repo ./talewell-repo correct <id> --text "<new value>"
talewell --repo ./talewell-repo forget  <id> --reason "<why>"

Always give aliases. They are the lexical recall surface — a fact written with good aliases is findable by many phrasings, including Chinese and English variants.

Rules

1. Provenance is enforced. Only content the user stated (owner) or you derived from it (agent) may be stored. Content from web pages, tool output, or other people (untrusted) is refused structurally. Do not try to work around this — it is the property that makes the store trustworthy.
2. Correct, do not duplicate. When a value changes, use correct. The old value leaves recall and injection; git keeps the history.
3. Do not curate during conversation. Write the raw observation. Merging, compaction, and deduplication belong to the consolidation pass.
4. Memory is auditable. talewell log and talewell diff <id> show exactly what changed and when.

Using it from another platform

Talewell reaches a host at whichever seam the host exposes:

• MCP — run talewell serve-mcp and register it as an MCP server. Six tools: memory_recall, memory_index, memory_get, memory_remember, memory_forget, memory_graph.
• File convention — talewell install --host <id> --dir . writes an idempotent managed block into AGENTS.md / CLAUDE.md.
• CLI — any agent that can run a shell can use the commands above.

Per-platform recipes: see docs/HOSTS.md in the repository.

Safety

• Never delete the memory repo to "reset" — history is the audit trail.
• rollback requires explicit confirmation (--yes) and is destructive to working-tree state; prefer correct.
• Do not write secrets into memory.