Two decisions separate this from a save-and-recall wrapper. Memories are classified as they are written and raw session logs are parked in a second tier, so recall does not hand back the transcript of the conversation you are already in. And facts get their own key-value store with a confidence ladder, because "what is my name" is a lookup rather than a search. It is one maintainer's production system rather than a hosted service, so you run the server, the model and the storage.
A local memory server with an MCP bridge in front of it. A Python API server holds memories in SQLite with FTS5 and semantic recall, and `integrations/mcp-bridge/server.js` is the stdio process your client actually launches. Memory is sorted at save time into two tiers: Tier 1 smart notes, classified into categories such as topology, current_state, doctrine, incident, identity, relationship and decision, and Tier 2 raw session logs tagged `session_log`, kept in full but excluded from default recall — pass `exclude_categories=[]` to search both. Several agents can point at one server, which is what makes cross-agent recall and overnight synthesis possible.
- Core memory over MCP — `mnemo_save`, `mnemo_recall`, `mnemo_search`, `mnemo_share`
- A structured fact store for what semantic search answers badly — `mnemo_fact_save`, `mnemo_fact_get`, `mnemo_fact_query`, `mnemo_fact_demote`, holding (entity, attribute, value) triples on a three-state confidence ladder: `verified` → `high_probability` → `false`, with contradictions raised over the bus and a Discord webhook
- Wiki pages still searchable through the bridge — `wiki_search`, `wiki_read`, `wiki_index`
- A hand-written project pad that registers only when `BRAIN_DIR` points at one — `read_brain_file`, `write_brain_file`, `list_brain_files`, plus `opie_startup` and `session_end`
- A beta behavioural-claim layer with a review queue, where nothing reaches the profile without an explicit promotion — `passport_get_user_context`, `passport_observe_behavior`, `passport_list_pending_observations`, `passport_promote_observation`, `passport_forget_or_override`
- Capture without touching the agent: `mnemo-cortex watch --backfill` tails Claude Code and OpenClaw session files from the outside, and `MNEMO_AUTO_CAPTURE=true` starts the watcher whenever the server starts
- Overnight cross-agent synthesis, so each agent begins the day knowing what the others did
- USB courier sync between two full installations — `mnemo-cortex stick init --encrypt`, `stick sync`, `stick watch --notify`, with optional AES-256 whose key never lives on the stick and conflicts preserved rather than overwritten
- A deployment check that names what failed — `mnemo-cortex health`, with `--json`, `--quiet`, `--agents`, `--services` and `--check-mcp <path>` for cron
- A bridge-level log of every tool call, off unless `MNEMO_DUMP=on` — one JSONL file per agent per day carrying `tool`, `params`, `response`, `latency_ms`, `ok` and `error`
Python 3.11 or newer for the server, and Node.js 18 or newer when you run the MCP bridge — that is the path Claude Desktop, LM Studio and OpenClaw use. Install from a clone with `pip install -e .`, which registers `mnemo-cortex` and the alias `mnemo`; `mnemo-cortex init` writes the config to `~/.agentb/agentb.yaml`, and `mnemo-cortex start` listens on `http://localhost:50001` with no auth token on loopback by default. Ollama is the recommended local reasoning and embedding provider, and the wizard also accepts OpenAI, Google, Anthropic or OpenRouter keys. The bridge reads `MNEMO_URL`, `MNEMO_AGENT_ID` and `MNEMO_SHARE`; where the server wants a key it comes from `MNEMO_AUTH_TOKEN` or `~/.mnemo-auth-token`. Linux, macOS and Windows are all supported — since v4.4.1 the server runs natively on Windows with no WSL. ChatGPT is not an MCP client here: it reaches the server through a bearer-authenticated, tenant-pinned two-route gate instead.