GitHub – StartripAI/codex-mem

Cut context tokens by 99.84% and reach first memory context in ~60 ms (median local benchmark).
Codex-native persistent memory with progressive retrieval, local viewer UX, and MCP-ready integration.

Quick Start • Comparison • Release Notes • Docs

Benchmark source:

Most coding assistants lose operational memory between sessions.
codex-mem makes new Codex sessions feel continuous by capturing lifecycle evidence, retrieving context progressively, and fusing memory with live repository facts.

North star:

• less repeated explanation
• less wasted context tokens
• more accurate follow-up reasoning from real prior work

This repository only keeps real product captures.
No placeholder GIFs or synthetic marketing screenshots are included.

• five hooks:
  • session-start
  • user-prompt-submit
  • post-tool-use
  • stop
  • session-end
• automatic session summary observations at close

• Layer 1: search / mem-search (compact shortlist)
• Layer 2: timeline (temporal neighborhood)
• Layer 3: get-observations (full details by selected IDs)

• ask combines memory shortlist with code context from repo_knowledge.py

• built-in natural-language search (nl-search / mem-search)
• local web viewer (stream, summary, search, config)
• runtime channel config (stable / beta)
• endless-mode style auto-compaction in beta
• dual-tag privacy model (--tag, --privacy-tag)

• MCP server with mem_* tools
• skill package for reusable retrieval workflow
• CLI wrapper for repeatable operations

See full history in RELEASE_NOTES.md.

Highlights in v0.3.0:

• launch asset production toolkit (Assets/LaunchKit/)
• CI asset gate (.github/workflows/asset-gate.yml)
• mem export surface (export-session, mem_export_session)
• web viewer upgrades (PRD caption copy + recording guide mode)
• launch automation scripts (make_gifs, validate_assets, social_pack, snapshot_docs)
• benchmark + roadmap + compatibility + security documentation set

Every release batch follows one fixed package:

• Release Notes update
• 3 GIFs (gif_01, gif_02, gif_03)
• 3 final screenshots
• 1 comparison table update in README

Reference:

bash Scripts/codex_mem.sh init --project demo

bash Scripts/codex_mem.sh session-start s1 --project demo --title "Streaming refactor"
bash Scripts/codex_mem.sh prompt s1 "Map end-to-end generation and persistence path" --project demo
bash Scripts/codex_mem.sh tool s1 shell "rg -n 'HomeStreamOrchestrator'" --project demo --title "Locate orchestrator" --compact
bash Scripts/codex_mem.sh stop s1 --project demo --content "checkpoint"
bash Scripts/codex_mem.sh session-end s1 --project demo

# Layer 1 compact search
bash Scripts/codex_mem.sh search "orchestrator streaming" --project demo --limit 20

# Layer 1 natural-language search
bash Scripts/codex_mem.sh mem-search "what bugs were fixed this week" --project demo --limit 20

# Layer 2 timeline
bash Scripts/codex_mem.sh timeline E12 --before 5 --after 5

# Layer 3 full details
bash Scripts/codex_mem.sh get E12 O3

bash Scripts/codex_mem.sh ask "What is the current generation chain from input to persisted output?" --project demo

Start:

bash Scripts/codex_mem.sh web --project-default demo --host 127.0.0.1 --port 37777

Open:

Viewer panels:

• real-time memory stream
• session summaries
• NL mem-search results
• runtime mode controls (stable / beta, refresh interval, endless mode)

post-tool-use supports two tag lanes:

• semantic tags: --tag
• privacy policy tags: --privacy-tag

Policy behavior:

• block write:
  • no_mem, block, skip, secret_block
• private visibility:
  • private, sensitive, secret
• redact sensitive values:
  • redact, mask, sensitive, secret

Example:

bash Scripts/codex_mem.sh tool s1 shell "credential=" --project demo \
  --tag auth \
  --privacy-tag private \
  --privacy-tag redact

Default retrieval hides private records unless --include-private is passed.

Read:

bash Scripts/codex_mem.sh config-get

Set:

bash Scripts/codex_mem.sh config-set --channel beta --viewer-refresh-sec 2 --beta-endless-mode on

Mode behavior:

• stable: compaction only when explicitly requested (--compact)
• beta + endless mode on: auto-compaction for high-volume tool outputs

Run:

python3 Scripts/codex_mem_mcp.py --root . --project-default demo

Register with Codex:

codex mcp add codex-mem -- python3 /ABS/PATH/codex-mem/Scripts/codex_mem_mcp.py --root /ABS/PATH/codex-mem --project-default demo

Retrieval:

• mem_search
• mem_nl_search
• mem_timeline
• mem_get_observations
• mem_ask

Runtime config:

• mem_config_get
• mem_config_set

Lifecycle:

• mem_session_start
• mem_user_prompt_submit
• mem_post_tool_use
• mem_stop
• mem_session_end
• mem_summarize_session
• mem_export_session

• Scripts/codex_mem.py core engine and CLI
• Scripts/codex_mem_mcp.py MCP server
• Scripts/codex_mem_web.py local web app
• Scripts/codex_mem_smoketest.py end-to-end simulation
• Scripts/repo_knowledge.py repository context retrieval
• Scripts/make_gifs.sh media pipeline (source -> webm/gif/poster)
• Scripts/validate_assets.py asset + README gate checks
• Scripts/load_demo_data.py one-click demo dataset loader
• Scripts/redact_screenshot.py OCR-based screenshot redaction
• Scripts/generate_social_pack.py X/Reddit/Product Hunt copy pack generator
• Scripts/compare_search_modes.py search vs mem-search comparison runner
• Scripts/snapshot_docs.sh release snapshot utility
• Skills/codex-mem/ skill package
• Documentation/ deep operational docs

Run one command:

python3 Scripts/codex_mem_smoketest.py --root .

Smoke test verifies:

• lifecycle capture
• NL mem-search
• privacy block/private/redact handling
• stable/beta config updates
• web APIs
• MCP tool registration and calls

# 1) load sanitized recording dataset
bash Scripts/load_demo_data.sh --reset

# 2) render GIF bundle from source clips
bash Scripts/make_gifs.sh --fps 12 --width 1200

# 3) validate media + README links
python Scripts/validate_assets.py --check-readme --strict

# 4) snapshot docs/media per release
bash Scripts/snapshot_docs.sh v0.3.0

# 5) generate social copy pack
python Scripts/generate_social_pack.py --version v0.3.0

Token reduction comes from retrieval discipline:

• compact shortlist first
• timeline around selected IDs only
• full payload fetch only for shortlisted IDs
• bounded repo context for fused ask
• optional/automatic output compaction for verbose tool logs

Primary channels:

• GitHub repo + releases
• MCP + Skills setup snippets in docs

Additional channels beyond GitHub/X:

• Product Hunt
• Hacker News (Show HN)
• Dev.to, Medium/Substack technical posts
• Reddit engineering communities
• Discord/Slack AI engineering communities

1. Branch from codex/init
2. Keep changes local-first and deterministic
3. Update docs with runnable examples
4. Run python3 Scripts/codex_mem_smoketest.py --root .
5. Include smoke output summary in PR

No explicit open-source license file is included yet.
Add one before broad redistribution.