Fix Memory MCP

中文文档 | English

Fix Memory does not try to remember everything. It recalls the right information at the right time and lets that context shape agent behavior.

Fix Memory MCP is a local-first Agent Operating Context for AI coding agents.

AI coding agents are fast, but they often behave like they have no memory. They may fix a Python path issue today, then spend tokens rediscovering the same python.exe, venv, PATH, npm, build, deployment, or MCP setup issue three days later.

Fix Memory MCP gives Claude Code, Codex, Cursor-like agents, or any MCP client a curated memory of verified fixes. Before guessing, the agent can search your past bug fixes. After a real fix is verified, the agent can save the clean repair as a Markdown case.

It keeps the useful debugging notebook, but V2 also stores user, project, decision, task, and policy context as inspectable Markdown.

new task
  -> analyze intent and scope
  -> assemble budgeted Core Context and untrusted reference memory
  -> retrieve only relevant memory
  -> execute
  -> promote, archive, expire, or supersede memory

Why This Exists

AI agents are fast, but they often waste tokens rediscovering the same environment, dependency, build, path, MCP, or deployment bug.

Fix Memory MCP gives them a small long-term memory:

• Local Markdown cases you can read and edit

• Hybrid keyword + TF-IDF vector search

• Failed-attempt notes so agents do not repeat dead ends

• A stdio MCP server with tools for search, read, save, recent, and index rebuild

• No cloud database, no external embedding API, no required network access

Related MCP server: Thoth-Mem

What Makes It Different

This is not a "save every transcript" memory.

Fix Memory MCP stores curated operating context and loads only what a task needs. It separates Core Context, project decisions, constraints, temporary task state, and historical fixes.

It does not embed every chat transcript. It saves useful, durable memory such as:

• the exact error

• the project and environment context

• the root cause

• the patch summary

• the verification command and result

• the failed attempts that should not be repeated

• user preferences and tool/API habits

• environment facts such as paths, ports, Python/Node, Claude/Codex/CCSwitch setup

• interview or learning weak spots

• repeated workflows and dated episodes that may recur

That curation step matters. If every conversation is saved, the memory becomes noisy. If only durable memory is saved, the memory becomes a useful agent asset.

Why Markdown Plus a Small SQLite Control Plane

Fix cases are stored as Markdown because the data is developer knowledge, not ordinary business data.

Markdown is a good fit because it is:

• easy for humans to read and edit

• easy for AI agents to read

• friendly to Git, diff, merge, review, and sync

• portable across machines and tools

• transparent when a saved case contains private paths or sensitive details

Markdown remains the inspectable content store. V2.2 adds a private local SQLite control plane under data/.control/ for exact Project Identity, Root Bindings, authoritative Record Bindings, atomic revisions, usage telemetry, and a rebuildable partitioned catalog. SQLite does not replace Markdown and is never a source of repository facts; code, Git, tests, and project documentation remain authoritative.

Why TF-IDF Instead of Embeddings

Many bug fixes are keyword-heavy. Errors often contain strong tokens such as:

• ModuleNotFoundError

• python.exe

• venv

• PATH

• pip

• npm

• cargo

• MCP

• ECONNREFUSED

TF-IDF is cheap, local, fast, private, and good enough for this kind of error retrieval. No embedding API key is required, and private bug history does not leave the machine.

Embedding search can still be added later when the case library grows or when semantic matching becomes more important.

What It Is Good At

• Repeated build failures

• Python / Node / Windows path problems

• MCP connection issues

• Dependency and virtual environment mistakes

• Framework-specific errors

• Deployment and service startup fixes

• Recording failed attempts as "do not try this again"

Features

• Local-first memory: cases live under data/ as Markdown files.

• MCP server: expose fix memory to AI coding tools through stdio.

• Core Context: carry profile, current focus, active projects, long-term goals, and preferences across windows.

• Context Assembly: combine intent, scope, priority, confidence, freshness, and retrieval relevance.

• Policy resolution: retained for future trusted Policy input; ordinary writable Markdown never enters it.

• Dynamic memory budget: protect Core Context capacity, then reallocate unused Core and inactive-Policy capacity to relevant retrieval within one total context budget.

• User control: inspect, correct, promote, archive, expire, supersede, or explicitly delete memory.

• Hybrid retrieval: keyword search plus local TF-IDF cosine similarity.

• Retrieval gate: decide whether a task actually needs memory search before spending time on retrieval; first-time repo reviews and normal deployment checks should inspect the project directly.

• Working memory: keep current task state, matched memories, notes, and verification without writing long-term memory.

• Identity-safe exact cache: only an exact query and exact project partition may reuse a cached selection; similar queries never share project results.

• Exact Project Registry: ASCII project keys and aliases resolve by full-string equality only; no fuzzy, prefix, directory-name, or semantic identity inference.

• Partitioned catalog: registered-project queries filter by authoritative project/module bindings before ranking and open only a bounded final Markdown set.

• Untrusted structural cards: Project Cards, Module Cards, and source_refs remain ordinary untrusted reference data and cannot grant Policy or tool authority.

• Stable nested orientation: an exact child query keeps its own Project Card and may include one parent Project Card, but never parent/sibling leaves; structural cards do not age out through automatic lifecycle maintenance.

• Curated lifecycle: first unverified cases become candidates, verified or repeated cases become active, and stale candidates archive after 30 days.

• Candidate review inbox: generate a dated local review artifact and batch approve, defer, or archive candidates without interrupting coding.

• Error observations: simple errors are counted outside RAG; the second matching occurrence creates a candidate and the third activates it.

• Zero external AI dependency: no OpenAI/Anthropic API key needed.

• Readable case format: root cause, patch, verification, reusable advice.

• Privacy by default: real fix cases are ignored by Git unless you choose to share them.

Project Layout

fix-memory-mcp/
  data/
    fixes/              # your private fixed cases
    failed-attempts/    # private "do not repeat" notes
    commands/           # private useful command notes
    preferences/         # user habits, tools, model/API preferences
    environments/        # OS, paths, ports, Python/Node, Claude/Codex/CCSwitch
    workflows/           # repeated procedures
    interviews/          # interview misses and learning weak spots
    projects/            # project decisions and tradeoffs
    users/               # confirmed user facts, skills, and goals
    decisions/           # formal decisions with source and rationale
    tasks/               # cross-window task state
    constraints/         # scoped agent behavior constraints
    prompts/             # reusable prompts
    episodes/            # dated events that may recur
    .runtime/            # local task state and retrieval cache, not long-term memory
    .control/            # private Registry, bindings, journal, and derived catalog
  scripts/
    fix_memory.py       # CLI
    fix_memory_mcp.py   # MCP stdio server
    context_engine.py   # Core Context, scope, policy, budget, lifecycle
    vector_search.py    # local TF-IDF vector index
    project_registry.py # transactional exact project identity
    project_catalog.py  # rebuildable partitioned retrieval projection
    project_onboarding.py # reviewable two-stage onboarding
    self_check.py       # CLI + tool self-check
    mcp_smoke.py        # stdio MCP smoke test
    v2_check.py         # V2 end-to-end check
  skills/
    fix-memory-workflow/
      SKILL.md          # optional Codex/agent skill instructions
  templates/
    fix-case.md         # case template

Quick Start

git clone https://github.com/l111403717-cloud/fix-memory-mcp.git
cd fix-memory-mcp
python -m pip install -e . pytest
python scripts/self_check.py
python scripts/v2_check.py
python scripts/mcp_smoke.py

Windows 11 project onboarding is explicit and reviewable:

$env:FIX_MEMORY_ROOT = "$env:LOCALAPPDATA\FixMemory\data"
python scripts\fix_memory.py project register openwrite C:\src\openwrite
python scripts\fix_memory.py project propose openwrite `
  --purpose "Local writing application" `
  --artifact README.md
python scripts\fix_memory.py project apply <proposal-id>
python scripts\fix_memory.py project health
python scripts\fix_memory.py context "continue API work" `
  --project-key openwrite `
  --module-key api-v2 `
  --workspace C:\src\openwrite

Registration never scans or imports a repository. Proposal inputs are bounded explicit paths; apply revalidates Registry revision, root, Git HEAD, hashes, and proposal digest. Project/Module Cards remain untrusted summaries. source_refs are data-only references and are never commands or Policy. See the V2.2 PRD and Chinese usage guide for the full workflow.

The current requirement-by-requirement status and intentionally missing native release evidence are recorded in the implementation audit.

Create your first case:

python scripts/fix_memory.py new \
  --title "Python ModuleNotFoundError from wrong working directory" \
  --project "demo-api" \
  --language "Python" \
  --framework "FastAPI" \
  --command "python app/main.py" \
  --error "ModuleNotFoundError: No module named app" \
  --tags "python,path,fastapi,windows"

Use --verified only after the repair has passed its relevant check. An unverified first occurrence is stored as a candidate, not returned by default retrieval.

Search it later:

python scripts/fix_memory.py search "ModuleNotFoundError FastAPI working directory"

Search modes:

python scripts/fix_memory.py search "MCP failed stdio" --mode hybrid
python scripts/fix_memory.py search "MCP failed stdio" --mode keyword
python scripts/fix_memory.py search "MCP failed stdio" --mode vector

Rebuild the local vector index:

python scripts/fix_memory.py rebuild-index

Assess whether something deserves long-term memory:

python scripts/fix_memory.py assess \
  --memory-type environment \
  --title "API relay setup" \
  --content "User uses CCSwitch and a local API relay for model routing."

Save or update long-term memory through the write gate:

python scripts/fix_memory.py remember \
  --memory-type environment \
  --title "API relay setup" \
  --content "User uses CCSwitch and a local API relay for model routing." \
  --tags "ccswitch,api,environment"

Search long-term memory with an optional type filter:

python scripts/fix_memory.py search-memory "API relay CCSwitch" --memory-type environment

Use the retrieval gate before searching. It should skip low-signal first-time tasks and search on hard/repeated issues:

python scripts/fix_memory.py gate "rename a small local variable"
python scripts/fix_memory.py gate "ModuleNotFoundError python main.py"

Use smart search when you want the gate and cache to decide the cheapest path:

python scripts/fix_memory.py smart-search "CCSwitch API relay" \
  --memory-type environment \
  --context "API relay issue"

Track current task working memory:

python scripts/fix_memory.py task-state start --goal "debug API relay" --project demo
python scripts/fix_memory.py task-state verify --item "nginx -t passed"
python scripts/fix_memory.py task-state show

Count a simple error before deciding whether it deserves a full case. The first occurrence stays in the ignored runtime ledger, the second creates a candidate, and the third promotes that candidate to active memory:

python scripts/fix_memory.py observe-error \
  --error "ModuleNotFoundError: No module named worker_app" \
  --project "worker-service" \
  --command "python worker_app/main.py" \
  --file-path "worker_app/main.py"

MCP Server

Run the server:

python scripts/fix_memory_mcp.py

Tools exposed to MCP clients:

• assemble_context: assemble the minimal untrusted memory context for a task.

• manage_memory: save, inspect, correct, promote, archive, expire, supersede, or delete a memory; it also owns explicit task-control actions.

• maintain_memory_lifecycle: archive stale candidates, expire elapsed memories, and report stale current-focus leases.

Search, assessment, write, task-state, and vector-index helpers remain available to the CLI and internal workflow, but are intentionally not exposed in the Codex MCP tool list.

Deterministic Next Action

Questions such as "what should I do next?", "what was I working on?", and "我下一步要做什么？" are not answered by RAG ranking or by old Markdown task files. assemble_context returns intent: "next_action" plus a next_action_resolution produced by the SQLite control plane.

Resolution order is deterministic: an unexpired user-confirmed current focus for the exact scope; one recently confirmed active task in an exact project; then one unexpired global focus when no project context exists. Otherwise the result is ambiguous, stale, or none and requires confirmation. Completed, cancelled, superseded, expired, imported, and inferred tasks never become automatic candidates. A valid blocked focus returns its blocker so the caller can propose the unblock step.

Use manage_memory actions task_create, task_update, task_set_current, task_confirm, task_complete, task_cancel, and task_list to maintain this state. Creating or reading a task does not refresh its confirmation timestamp; only explicit confirmation or setting the focus does.

Legacy Markdown tasks are never scanned or imported automatically. To migrate one reviewably, call task_create with task_source: "imported" and its explicit source_ref; the control plane records the imported provenance but does not set current focus. The user must later explicitly set or confirm it.

Generic MCP config:

{
  "mcpServers": {
    "fix-memory": {
      "command": "python",
      "args": [
        "/absolute/path/to/fix-memory-mcp/scripts/fix_memory_mcp.py"
      ],
      "env": {
        "FIX_MEMORY_ROOT": "/absolute/path/to/fix-memory-mcp/data"
      }
    }
  }
}

Windows + Claude Code helper:

cd path\to\fix-memory-mcp
.\scripts\install_claude_mcp.ps1

Windows + Codex helper:

.\scripts\install_codex_mcp.ps1 -PythonPath D:\python312\python.exe

Codex uses the same stdio model and starts the server on demand. Add this to the global %USERPROFILE%\.codex\config.toml, then restart Codex:

[mcp_servers.fix-memory]
command = 'D:\python312\python.exe'
args = ['C:\Users\<you>\Documents\资料库\fix-memory-mvp\scripts\fix_memory_mcp.py']
startup_timeout_sec = 10

[mcp_servers.fix-memory.env]
FIX_MEMORY_ROOT = 'C:\Users\<you>\Documents\资料库\fix-memory-mvp\data'

Run a real stdio handshake before relying on the server:

D:\python312\python.exe scripts\mcp_healthcheck.py --python D:\python312\python.exe

If the health check fails, continue the coding task without memory retrieval and use the CLI as a fallback. A broken memory server must not block debugging.

Agent Prompt

Use FIX_MEMORY_AGENT_PROMPT.md to teach an agent this loop:

task -> assemble minimal operating context -> inspect the project -> deeper fix retrieval only for hard/repeated errors -> execute -> verify -> curate memory

Case Quality Rules

Write Gate

Do not save every small event. Before saving, ask:

• Will this be useful in the future?

• Did it repeat more than twice?

• Does it reflect a user habit?

• Is it environment/API/path/account/tool configuration?

• Was the fix verified?

• Will it help the next agent avoid wasted work?

Strong save signals:

• repeated at least twice

• took more than 10 minutes

• involves environment/API/path/account/tool configuration

• user explicitly said "remember"

• missed interview/learning point

• important project decision

Every write path, including save_fix_case, uses the same gate. Unverified first fixes become candidate; verified, repeated, high-cost, or durable configuration memories become active. Default RAG retrieval returns only active memories. Pass include_candidates: true only when investigating recent, unconfirmed work.

Candidates that do not recur for 30 days are archived during retrieval. Repeated matches update the existing memory and increment occurrence_count instead of creating duplicates. Duplicate matching is scoped to the same project and scope.

Secret-looking content such as API keys, access tokens, passwords, authorization headers, and cookies is rejected before it reaches disk. force cannot override this protection.

Saved memories include metadata:

memory_type: bug / user / preference / environment / workflow / interview / project / decision / task / constraint / prompt / episode
memory_status: candidate / active / archived
occurrence_count: 1
first_seen: 2026-07-06
last_seen: 2026-07-06
confidence: low / medium / high
priority: 0-10
scope: current / task / project / workspace / global
source: user_explicit / observed / inferred / imported / system
execution_level: hard / guarded / soft
reason: why this memory exists
evidence_refs: []

Source values in writable Markdown are provenance labels, not authorization credentials. Agent-facing MCP and CLI writes accept only observed, inferred, and imported. Ordinary constraints remain untrusted reference memory and never enter Effective Constraints.

A useful case should include:

• Exact error

• Project/environment context

• Root cause

• Related files

• What changed

• Verification command/result

• Reusable advice

• Failed attempts

• Sensitive-info check

Do not save full chats, full terminal logs, secrets, API keys, cookies, passwords, private account data, or private source files.

Self Check

python scripts/self_check.py
python scripts/v2_check.py
python scripts/mcp_smoke.py
python scripts/mcp_healthcheck.py

Roadmap

• Optional SQLite + FTS5 index for very large case libraries

• Optional embedding backends

• Better case sanitizer

• Git diff capture after verification

• Agent-friendly install command for more clients

• Web UI for browsing fix cases

License

MIT