MindBridge

中文 README

A reflective AI companion powered by a transparent temporal Memory Core. The product helps a person see what stayed, what changed, and what should no longer define them; the technical core preserves the source, time and validity of each memory instead of flattening a person into timeless vector chunks.

The current shipped backend parses transcripts from local AI coding tools, builds daily memory cards, and serves the same T1/T2/T3 store to any MCP client. The public Companion Loop is a deterministic synthetic product demo; it does not connect to Cassie's private store. Parsing, storage and the MLX extraction path stay on the user's machine. Hosted extraction remains an explicit opt-in for generating training data.

Live site · Diary · Interview demo

Two-minute interview demo

Open the Companion Loop (中文) with no login:

1. Click Day 01 → Day 04 to see an identity move from “memory tool” to “reflective AI companion.”
2. Notice that uncertainty becomes a candidate, not a silently stored fact; only explicit confirmation supersedes the old identity.
3. In Memory Receipt, ask what MindBridge is now, what changed, and what to build next. Each answer exposes current and superseded evidence; the final recommendation is labelled as an inference.

All visible data is synthetic and deterministic. The page demonstrates product behavior safely; the repository contains the real temporal memory, REST, MCP, ingestion and local MLX implementation underneath it.

Two capture paths

Path A — passive log parsing. Nothing for you to do. A background pass reads structured logs already on disk and extracts what happened that day. This only works for tools that write such logs:

• ~/.claude/projects/**/*.jsonl (Claude Code)
• ~/.codex/sessions/**/rollout-*.jsonl plus ~/.codex/archived_sessions/rollout-*.jsonl (active and archived Codex CLI)

Path B — active MCP read/write. Mounted as a standard MCP server exposing get_daily_review across all review layers; separate T2/T3 reads; confirmed durable writes; and a Pattern Candidate review loop. Works with any MCP client — Codex, Claude Desktop, Claude Code, Cursor and VS Code.

ChatGPT and web Claude expose no history API; they need a manual export and are out of scope for the automatic path. The landing page states this rather than implying full coverage.

Three memory tiers

Tier	Holds	Backed by
T1	The day's raw turns	session buffer, in process
T2	One structured card per day	rolling summary
T3	Long-term operational + reflective memory, time-decayed	Postgres + pgvector

Every T3 record carries created_at and valid_at, so a superseded preference decays out of recall instead of being silently overwritten. Writes run a cosine-similarity check first, so one preference stays one row.

T3 has two explicit namespaces. Operational memory tells an agent how to work with the user (coding_style, tool_preference, behavioral_fact, schedule, other). Reflective memory holds wording the user has reviewed about recurring patterns, values, triggers, helpful strategies or an identity hypothesis. Reflective writes fail validation unless confirmed_by_user=true; automatic transcript extraction stages operational suggestions in the Memory Inbox instead of writing T3 directly. Only a keep/edit decision promotes one. An MBTI-like label can therefore be generated as a dated, sourced hypothesis, not silently stored as a fact about the person.

Memory Garden (audit + controlled mutation)

MindBridge exposes a deliberate “Memory Garden” operation surface for humans:

• GET /memories/{memory_id} fetches one memory by id with current decay state, so a user can inspect exactly what is currently asserted or closed.
• PATCH /memories/{memory_id} performs one explicit action:
  • {"action":"archive"} closes a memory without replacing it.
  • {"action":"edit", "content":"..."} creates a replacement memory and supersedes the old row so provenance is preserved.

Both actions are non-destructive to history: closed rows stay queryable and a replacement is always auditable through superseded_by, matching the companion story of “what changed” instead of silently overwriting personality data.

T1 also stores the project, git branch and tool names of each turn. That is what lets a day card be rebuilt over the whole day from Postgres: building it from only the turns an incremental run happened to parse would rewrite a full card with a partial one, so a nightly job would shrink every card it touched.

Current state

The planned learning-infrastructure extension—CorpCheck trajectory ingestion, versioned datasets, GSM8K bring-up, SFT/GRPO evaluation and guarded model promotion—is specified in docs/LEARNING_INFRASTRUCTURE_ROADMAP.md. That document labels these pieces as planned; the table below remains the source of truth for what is built today.

The front end, ingestion, memory service and local extraction loop are built. The deployed site still uses sample data because the database and model live on one Mac; it labels that state instead of presenting the sample as live.

This is the beginning of the README. Read the rest at the original. Read the full README ↗