codex-issue-memory MCP server

codex-issue-memory

codex-issue-memory is a local-first MCP server for Codex. It stores verified fixes, retrieval feedback, preferences, and guardrails in SQLite so repeated failures can be diagnosed from prior evidence instead of re-debugged from scratch.

What it provides

• MCP runtime: python -m codex_issue_memory.server
• Public MCP tools: 12 tools for matching, inspection, feedback, preferences, guardrails, metrics, and review
• Maintenance CLI: issue-memory-maint for schema setup, health checks, backups, telemetry, benchmarks, and lifecycle diagnostics
• Local persistence: SQLite DB plus state/log/backup directories under configurable local paths
• Codex integration: one authoritative live registration in ~/.codex/config.toml
• Optional wrapper metadata: .codex-plugin/plugin.json and .mcp.json for local/custom plugin workflows only

When to use it

Use this project when you want to:

• search prior fixes from a short error excerpt,
• keep reusable debugging knowledge local,
• record only verified reusable resolutions,
• learn from accepted/rejected retrievals,
• inspect server health, backups, and lifecycle state,
• keep Codex MCP setup reproducible across sessions.

Quick start

git clone https://github.com/PhiniteLab/codex-issue-memory.git
cd codex-issue-memory
bash install.sh
bash scripts/verify_install.sh

The installer prepares a virtualenv, initializes the SQLite store, writes the live issue_memory MCP block to ~/.codex/config.toml, and runs a local verification pass.

Installation modes

Install Mode A — Standard MCP install

Use this for the real runtime.

bash install.sh
bash scripts/verify_install.sh

Verify the live registration:

grep -n '^\[mcp_servers.issue_memory\]' ~/.codex/config.toml
issue-memory-maint smoke
issue-memory-maint doctor --mode shadow --max-instances 0

Expected posture:

• exactly one live [mcp_servers.issue_memory] block,
• writable Linux/WSL-local SQLite/state paths,
• successful smoke and doctor output.

Manual package setup

python3 -m venv .venv
. .venv/bin/activate
python -m pip install --upgrade pip setuptools wheel
python -m pip install -e .[dev]
python -m codex_issue_memory.maintenance init-db
python -m codex_issue_memory.server

Then add one live [mcp_servers.issue_memory] block to ~/.codex/config.toml.

Install Mode B — Plugin-wrapper metadata

Use this only when you need local/custom plugin metadata or manual marketplace-style integration.

python3 -m json.tool .codex-plugin/plugin.json
python3 -m json.tool .mcp.json

Mode B does not create the Python runtime, initialize the database, or register the live MCP server. For runtime use, choose Mode A or manual registration.

MCP tools

Retrieval and inspection

• issue_match — find likely prior issue patterns for a failure
• issue_get — load full details for one pattern
• issue_search — keyword search stored patterns
• issue_recent — list recent patterns

Write-back and learning

• issue_record_resolution — store a verified reusable fix
• issue_feedback — record accepted/rejected/fix feedback

Preferences and guardrails

• issue_set_preference — save prompt-driven strategy preferences
• issue_list_preferences — list saved preferences
• issue_guardrails — retrieve proactive prevention guidance

Operations and review

• issue_metrics — report operational metrics
• issue_review_queue — list pending/resolved review items
• issue_review_resolve — resolve review queue items

Maintenance CLI

Common commands:

issue-memory-maint smoke
issue-memory-maint smoke-learning
issue-memory-maint doctor --mode shadow --max-instances 0
issue-memory-maint server-status
issue-memory-maint metrics --window-days 30
issue-memory-maint review-queue --status pending --limit 20
issue-memory-maint e2e-mcp-reuse-harness --json

Useful command groups:

• Schema: init-db, migrate-v2, schema-version
• Backups: backup, list-backups, verify-backup, restore-backup
• Health/lifecycle: smoke, smoke-learning, server-status, runtime-diagnostics, recommended-config, doctor, e2e-mcp-reuse-harness
• Telemetry/retention: metrics, export-dashboard, prune-retention
• Review queue: review-queue, resolve-review
• Benchmarks/calibration: quality-gate-benchmarks, benchmark-user-domains, benchmark-failure-taxonomy, benchmark-dense-bandit, benchmark-real-world, benchmark-hard-negatives, benchmark-merge-stress, calibrate-thresholds, calibrate-weights, analyze-feature-importance, sweep-implicit
• Experiment registry: create-experiment, update-experiment, analyze-experiment

For the complete CLI surface:

issue-memory-maint --help

Python API example

from codex_issue_memory.app import IssueMemoryApp

app = IssueMemoryApp()

result = app.issue_match(
    error_text="ModuleNotFoundError: No module named requests",
    command="python worker.py",
    file_path="api/worker.py",
    project_scope="my-repo",
)
print(result["decision"])

app.issue_record_resolution(
    title="Missing dependency in worker runtime",
    raw_error="ModuleNotFoundError: No module named requests",
    canonical_fix="Install the dependency in the same runtime environment used by the target process.",
    prevention_rule="Pin and install runtime dependencies before process startup.",
    project_scope="my-repo",
)

Architecture flow

Failure excerpt
  ↓
Normalize into a query profile
  ↓
Retrieve candidates from SQLite-backed memory
  ↓
Rank and decide: match / ambiguous / abstain
  ↓
Return compact guidance through MCP
  ↓
Record feedback and verified reusable fixes

Runtime authority and safety

• Keep the live MCP authority in ~/.codex/config.toml.
• Keep exactly one [mcp_servers.issue_memory] block unless you intentionally manage alternatives.
• Prefer Linux/WSL-local writable paths for DB, state, logs, and backups.
• Treat .mcp.json, .codex-plugin/, templates/, and bundled skills/ as distribution/reference material, not the live runtime authority.
• Write memories only after a fix is verified and reusable.
• Use project-specific scope when a memory should not be global.

Validation

Recommended local checks:

bash scripts/verify_install.sh
python -m pytest
python -m codex_issue_memory.maintenance smoke
python -m codex_issue_memory.maintenance server-status
python -m codex_issue_memory.maintenance e2e-mcp-reuse-harness --json

Optional checks when available:

pyright
ruff check src tests scripts

Troubleshooting

• Codex does not show issue_memory: restart Codex and verify exactly one live [mcp_servers.issue_memory] block.
• Install problems: retry with SKIP_CRON_INSTALL=1 bash install.sh and confirm Python/venv access.
• Weak retrieval quality: pass a short meaningful error excerpt plus command, file path, and project scope.
• DB path or permission issues: keep the writable DB under local Linux/WSL storage such as ~/.local/share/codex-issue-memory.

Documentation map

Start with docs/README.md. Key docs:

• docs/INSTALLATION.md
• docs/CONFIGURATION.md
• docs/USAGE.md
• docs/OPERATIONS.md
• docs/OWNER_KEY_CONTRACT.md
• docs/ORCHESTRATION_STDIO_REUSE_CHECKLIST.md
• docs/ARCHITECTURE.md
• docs/DEVELOPMENT.md
• docs/ROADMAP.md

Repository layout

src/codex_issue_memory/             MCP runtime, services, storage, matching, maintenance
src/codex_issue_memory/sql/         SQLite migrations packaged with the runtime
scripts/                            install, registration, backup, and verification helpers
docs/                               public and contributor documentation
templates/                          reference snippets and plugin-wrapper examples
skills/                             bundled reference skill content
tests/                              regression, lifecycle, diagnostics, and benchmark tests
reports_lifecycle_investigation/    historical lifecycle debugging notes and test evidence
.codex-plugin/                      local/custom plugin metadata
.mcp.json                           wrapper server templates

Contributing

Contributions improving install clarity, retrieval quality, operational safety, and validation are welcome. See CONTRIBUTING.md and docs/DEVELOPMENT.md.

License

MIT. See LICENSE.