Memory

Memory is one shared set of facts across all your agents. A fact written by any agent (or by you) is stored as a Markdown file under ~/.coffer/memory/, recalled by every other agent through Coffer's gateway, and projected natively into each agent's own memory location so no per-agent copy drifts.

Memory shares the knowledge substrate: files are the source of truth, the SQLite index is rebuildable, and the same three retrieval modes apply.

Two scopes

• global — applies everywhere (~/.coffer/memory/global/).
• per-project — scoped to one project, resolved from the git root of the agent's working directory.

Stores are auto-provisioned per scope; you never "create" one. recall returns both scopes by default.

Add and recall

coffer memory add global "Prefer pnpm over npm in all repos" --name pkg-manager
coffer memory recall global "which package manager?"     # spans project + global
coffer memory facts global                               # list facts
coffer memory edit global <fact-id> "…"                  # curate
coffer memory delete global <fact-id>

• Facts are stored verbatim — there is no LLM at write time.
• recall rescans the fact directory for out-of-band edits before searching, and uses the same grep / keyword / vector modes as a knowledge base.

Native projection

Binding a store to an agent projects its facts into that agent's own memory:

coffer memory bind <store> claude-code --project-root /path/to/repo
coffer memory unbind <store> claude-code

• Claude Code — the project memory directory is symlinked into ~/.claude/projects/<slug>/memory/, and the auto-generated MEMORY.md index follows Claude Code's own memory format.
• Codex — facts render into a managed block in AGENTS.md; Codex's native memory is disabled so there is never a second copy.

Agents also get MCP tools: coffer__remember, coffer__recall, coffer__update_memory, coffer__forget, and coffer__list_memory. A desktop UI lets you browse and curate facts per scope.

Channels →