Memory (cavemem)

Me Write Code delegates persistent memory to cavemem. Me Write Code's value-add is policy: when to write, what to inject, and the episodic→semantic consolidation pass.

What cavemem provides

• SQLite (FTS5) storage with hybrid search: BM25 + local vectors (Xenova/all-MiniLM-L6-v2, alpha=0.5).
• Prose-token reduction via compact grammar. Code, paths, and URLs preserved byte-for-byte.
• Stdio MCP server with four tools: search, timeline, get_observations, list_sessions.
• Hook CLI: cavemem hook run <event> reads stdin and writes observations.
• Privacy: <private>...</private> blocks are redacted before storage.

How mewrite-code uses it

mewrite session
  ├── on session_start  → cavemem hook run session-start  (write)
  ├── on user_prompt    → cavemem hook run user-prompt-submit  (write)
  ├── on post_tool_use  → cavemem hook run post-tool-use  (write, async)
  ├── on stop           → cavemem hook run stop  (write)
  └── reads via cavemem MCP (search, timeline, observations, sessions)

On session start, mewrite-code runs cavemem search "<task summary>" and injects compact snippets into context. The injection is capped at 2k tokens by default (memory.maxInjectTokens).

Setup

Cavemem is the default memory backend. Me Write Code probes cavemem --version and uses it when available:

npm install -g cavemem
mewrite mcp add cavemem --command cavemem --arg mcp

Memory recall/tools are enabled when the provider is reachable. To auto-record lifecycle observations through hooks, start sessions with CAVE_MEMORY_AUTO_RECORD=1 or wire the cavemem hook commands manually in ~/.mewrite/agent/settings.json.

To disable for a session: /memory off.

Slash commands

Command	What it does
/memory search <query>	Hybrid search across all sessions
/memory save <text>	Write an explicit observation (kind: explicit)
/memory show <id>	Expand a snippet to full body
/memory forget <id>	Soft-delete an observation
/memory export [--format md|json]	Dump memory
/memory consolidate	Me Write Code-specific: cluster recent observations, ask Haiku for semantic facts, write back as kind:semantic
/memory off /memory on	Pause/resume injection for the current session
/memory config	Edit memory settings
/memory sync --from claude	One-shot import of ~/.claude/projects/<slug>/memory/MEMORY.md

Me Write Code's value-add

Episodic→semantic consolidation

Run nightly (via cron) or on-demand with /memory consolidate. Me Write Code clusters observations by topic, asks Haiku to extract semantic facts, writes them back as kind: semantic with provenance ids pointing at the source episodic observations. This closes a loop most agents skip — what makes Letta and Zep feel "smart" — but local, deterministic, and cheap.

# nightly cron
0 2 * * * mewrite memory consolidate --since 24h --model haiku

Auto-trigger learning

When a tool call fails twice and then succeeds, mewrite-code writes a "lesson" observation:

kind: lesson
context: "applying Biome config to a TypeScript monorepo"
fail: "biome.json with deprecated lint key"
fix: "rename to linter, drop the legacy formatter block"
provenance: [obs_id_1, obs_id_2, obs_id_3]

Mirrors Claude Code's Auto-Memory.

MEMORY.md bridge

On session start, mewrite-code reads ~/.claude/projects/<slug>/memory/MEMORY.md (first 200 lines) so it behaves consistently when invoked in a project where Claude Code is also active.

mewrite memory sync --from claude

Imports the per-fact .md files as cavemem observations.

Privacy

• Anything between <private> and </private> is dropped before write. Use it for credentials, names, etc.
• Me Write Code never sends memory content to a model unless explicitly injected (search results, get_observations).
• Storage is local: ~/.cavemem/. No telemetry, no cloud.

Falling back to files

When Cavemem is unavailable, Me Write Code automatically uses plain .mewrite/memory/*.md files and project context docs. To use FilesProvider without probing Cavemem, set memory.backend: "files" in ~/.mewrite/agent/settings.json.

{
  "memory": {
    "backend": "files"
  }
}

Troubleshooting

• cavemem: command not found — install it (npm i -g cavemem) to use Cavemem; otherwise Me Write Code falls back to FilesProvider. Set memory.backend: "files" to skip the probe.
• Memory injection too aggressive — lower memory.maxInjectTokens.
• Want to wipe — rm -rf ~/.cavemem/.