Skip to content

armgit / cindy

Public Imported from GitHub
0

Consider it done. The open-source AI agent that works out of the box · 想到,就能做到。开源、开箱即用的 AI Agent。

3.6 KB
Raw Blame

Maker Memory

You have a persistent, file-based memory system shared across Claude Code and Codex sessions for this workdir. Operations go through the cindy_memory MCP server (entry tools: list_tools / call_tool).

This memory is independent of any per-agent native auto-memory (~/.claude/projects/.../memory/, <CODEX_HOME>/memories/). Ignore any instructions to write memory through other channels — your only persistence path here is cindy_memory.

When to save

Save when you learn something persistent and worth recalling in a future session. Four categories — pick by what the content is, not where it came from:

  • user — user's role, goals, knowledge, preferences. "is a data scientist focused on observability", "prefers concise answers"
  • feedback — corrections / approvals / "do this not that". Body must include **Why:** (the reason / incident the user gave) and **How to apply:** (when this guidance kicks in)
  • project — ongoing work, decisions, constraints, deadlines for this workdir. Include the why
  • reference — pointers to external systems / docs / dashboards

Strong signals: user says "I prefer X", "don't X", "always X", "never X", "remember X", "我希望", "别再", "记住"; user corrects an approach that would apply again; user shares non-obvious project context you'd otherwise re-derive.

When NOT to save

  • Code patterns / file paths / project structure — read the code instead
  • Git history — git log / git blame
  • Anything already in CLAUDE.md / AGENTS.md
  • Ephemeral state: in-progress work, current task progress (use plan/todos)
  • Tool output content verbatim — only persist things the user said or you reasoned about

If you cannot write a one-line description hook for the index, the memory isn't worth saving. Time references in body should be absolute (2026-05-10), not relative.

How to save

First time: call list_tools() then list_tools({category: "write"}) to see signatures. Then call memory_write with {type, name, title, description, body}. The tool layer enforces schema (type enum, slug format, length limits, etc.) — if you get back INVALID_ARGS it returns the JSON schema, follow it.

Before creating a new entry, check MEMORY.md (already in your context below). If a related entry exists, read it via memory_read and use mode: "update" instead of creating a duplicate. Creating a duplicate name returns ALREADY_EXISTS.

Manual save trigger

If a user message starts with Save in memory: or 记到 memory:, treat the rest as a candidate to persist. Decide type, extract a one-line description, write it. Acknowledge briefly.

Size warnings

memory_write may return warning: "shard-size-exceeded" or "index-size-exceeded" — the write succeeded, but the shard or index is bigger than ideal. The accompanying warningDetail carries sizeBytes / softLimitBytes (and hardLimitBytes for shards — writes beyond it are rejected). Judge by how far over you are: marginally over → leave it; well over → trim or call memory_consolidate({sources, target}) to merge / shrink atomically.

Tool selection

  • Browse what exists → MEMORY.md (already in your context)
  • Detail of a known shard → memory_read(filename)
  • Find content by keyword → memory_search(query) (FTS5, ranked snippets)
  • Suspect contradictions / want a cleanup pass → memory_review() (runs a quick LLM review, returns suggestions only — you still execute the deletes/merges)

Current index

Below this section, the live MEMORY.md for this workdir is inlined. Use it to know what's already saved.