Lossless Context Management for Recursive Language Model
Project description
MemoryForge
MemoryForge is a Codex-CLI-first, local-first memory layer for long-running LLM workflows.
It stores durable project evidence in SQLite, keeps live context bounded, and returns source-backed context bundles to the model or agent the user is already running. For Codex CLI usage on Linux/WSL, the project-local hook runner is the required LCM auto-capture path. Codex MCP remains deliberately small and exposes only hot-path recall/context plus the RLM analysis planner.
Codex CLI -> WSL/Linux hook runner -> same SQLite memory.db
Codex CLI -> MemoryForge MCP -> recall/context/index_analyze
Codex host subagents -> fetch chunks -> record -> aggregate
What It Is For
MemoryForge focuses on long-form project evidence that is common in real coding workflows but too large or noisy to paste into every prompt:
- design notes and architecture documents
- requirements and implementation plans
- setup guides and decision records
- benchmark descriptions and experiment logs
- large Markdown files used while "vibe coding" or building a project over many sessions
These sources are ingested through RLM, stored as durable LTM evidence, and
recalled as bounded context when Codex needs them. The shallow path is the RLM
sub-agent analysis/summary stored in LTM; the deep path remains the lossless
rlm_chunk:<id> content that can be rehydrated on demand.
MemoryForge intentionally avoids a large AST/code graph schema in the core
release. The default auto-load path targets Markdown project knowledge
(README, design notes, ADRs, plans, reports, specs). Code files can still be
ingested manually, but dedicated code indexing is future work rather than part
of the current SQLite schema.
Memory Layers
| Layer | Role | Model worker use |
|---|---|---|
| RLM | Raw Large Memory. Chunks large files/prompts, prepares host-subagent analysis plans, and indexes both derived summaries and full chunks into durable memory. | Host agent/subagent executes the returned plan. |
| LTM | Long-Term Memory. Recalls durable evidence across sessions and sources. | No model call. |
| LCM | Lossless Context Management. Keeps the MemoryForge active-session view bounded with summaries, raw refs, and recoverable tool-output parts. | WSL/Linux hook runner captures Codex CLI lifecycle; optional explicit compaction worker. |
Important boundary: LCM compacts MemoryForge's SQLite-backed active context. It
does not directly erase Codex's own context window. Codex manages its live
thread and can compact it with /compact; MemoryForge MCP tools and the
WSL/Linux hook runner preserve and rehydrate the evidence needed after
compaction.
Install
From PyPI in a project that uses uv:
uv add memfg==6.1.5
MemoryForge defaults to lexical BM25/FTS recall so first run stays responsive on
fresh PyPI installs and offline machines. Semantic/vector recall is available
when explicitly enabled with MEMORYFORGE_VECTOR_BACKEND=fastembed.
For local development from this repository:
uv sync --extra dev --extra benchmark
PyPI 6.1.5 Post-Publish Smoke Test
After publishing:
uvx --from twine twine upload dist/memfg-6.1.5*
use a clean project to verify the package from PyPI. The flow below is written for Windows PowerShell, but the same commands work in any shell after adapting path syntax.
- Create a fresh uv project outside this repository:
New-Item -ItemType Directory -Force C:\tmp\memoryforge-pypi-smoke | Out-Null
Set-Location C:\tmp\memoryforge-pypi-smoke
uv init --bare --name memoryforge-pypi-smoke .
- Install the freshly published PyPI package:
uv add memfg==6.1.5
uv run memoryforge --help
uv run memoryforge --help must print CLI usage and exit. Do not use
uv run memoryforge-mcp as a normal smoke test; that command starts the MCP
stdio server and waits for a client, so an idle terminal there is expected.
- Register the MCP server with Codex CLI:
codex mcp remove memoryforge
codex mcp add memoryforge -- uv run memoryforge-mcp
codex mcp get memoryforge --json
If codex mcp remove memoryforge says the server does not exist, continue with
the codex mcp add command. Restart Codex after changing MCP registration.
- Initialize MemoryForge in the project:
uv run memoryforge init . --agent-id codex --force
Expected behavior for 6.1.5:
- The command exits by itself.
.memoryforge\memory.dbis created..memoryforge\config.jsonis created.AGENTS.mdis created or updated.- The JSON output shows
"indexed": {"enabled": false, ...}. - The JSON output shows
"codex /init not requested"unless you passed--configure-codex.
init is intentionally lightweight. It does not index Markdown by default, and
it does not call Codex CLI subprocesses by default.
- Add a small Markdown memory file:
New-Item -ItemType Directory -Force docs | Out-Null
@'
# Facilities telemetry
Facilities use telemetry ingress endpoint https://telemetry.facilities.example.com/v2/ingest.
The old endpoint https://telemetry.old.example.com/ingest was rejected because it bypassed tenant isolation and failed TLS pinning.
'@ | Set-Content -Encoding UTF8 docs\telemetry.md
- Index the Markdown files. The default path is fast raw chunk/LTM indexing and does not call Codex sub-agents:
uv run memoryforge index . `
--agent-id codex `
--max-files 1 `
--force
For RLM sub-agent analysis, do not leave Codex to run a separate CLI command.
Start Codex and ask it to use MCP index_analyze; the returned
host_subagent_prompt values are the host-subagent tasks.
init --index remains accepted for compatibility, but the recommended command
is memoryforge index.
- Verify recall without Codex:
uv run memoryforge --db .memoryforge\memory.db recall-memory `
--agent-id codex `
--query "telemetry ingress endpoint facilities old value rejected" `
--include-content
The output should include https://telemetry.facilities.example.com/v2/ingest
and the rejection reason about tenant isolation and TLS pinning.
- Install and trust the Codex lifecycle hook from WSL/Linux if you want LCM to capture interactive Codex turns:
uv run memoryforge init . --agent-id codex --force --install-hooks
codex
Inside Codex, run /hooks and trust the MemoryForge project-local hooks. Without
this trust step, LCM capture is unavailable for the Codex session.
- Optional: verify recall and RLM planning through Codex MCP. Start Codex from the same project directory:
codex
Ask:
What is the telemetry ingress endpoint used by facilities, and why was the old value rejected?
Expected Codex behavior:
- It should call
memoryforge.recall_memory. - The tool call should return quickly for this small project.
- The answer should cite the new endpoint and explain why the old value was rejected.
If Codex does not call MemoryForge, ask explicitly:
Use MemoryForge MCP recall_memory first. What is the telemetry ingress endpoint used by facilities, and why was the old value rejected?
To make Codex prepare RLM host-subagent work from inside the active session, ask:
Use MemoryForge MCP index_analyze for this project with analyze_min_bytes=20000 and analyze_max_files=5, then dispatch the returned host_subagent_prompt batches.
- Verify LCM lifecycle capture for Codex interactive mode. Use WSL/Linux rather than native PowerShell hooks:
uv run memoryforge init . --agent-id codex --force --install-hooks
Expected files:
.codex/hooks.json
.memoryforge/hooks/memoryforge-hook.sh
.memoryforge/hooks/memoryforge-hook.log
Restart Codex from the same WSL/Linux project directory:
codex
Then run /hooks in Codex and trust the MemoryForge hook definitions. Codex
requires this review for project-local command hooks. The MemoryForge hook is
intentionally small: it calls python -m memoryforge.cli.main hook ... through
the project Python when available, uses uv run --no-sync as a fallback, and
has a 30 second timeout. It does not call codex, does not start a model
worker, does not use the Codex account/token, does not sync/reinstall the
project, and does not index the whole project on startup unless
MEMORYFORGE_HOOK_AUTO_INDEX=1 is set. Diagnostics go to
.memoryforge/hooks/memoryforge-hook.log.
If you previously installed hooks with an older MemoryForge build and Codex
shows SessionStart hook (failed) or UserPromptSubmit hook (failed), remove
the old native Windows hook files and regenerate from WSL/Linux:
uv run memoryforge init . --agent-id codex --force --install-hooks
Then reopen Codex from WSL/Linux and trust the hooks again with /hooks.
The hook path listens for:
SessionStart: cleans stale pending hook files. It does not auto-index Markdown by default.UserPromptSubmit: stores the pending user prompt and records an LCM context snapshot.PostToolUse: stores tool output as an assistant message with atoolpart when Codex supplies tool output in the hook payload.Stop: commits the completed turn; if Codex supplies assistant output, it stores user + tool + assistant, otherwise it still commits the user prompt so the session is not empty.PreCompactandPostCompact: record context snapshots around Codex/compact;PostCompactalso stores a compact summary if Codex supplies one.
Ask Codex one real question, then inspect the MemoryForge LCM database:
uv run memoryforge --db .memoryforge\memory.db lcm-sessions `
--agent-id codex
uv run memoryforge --db .memoryforge\memory.db lcm-messages `
--session-id <session-id-from-lcm-sessions> `
--agent-id codex `
--include-content
uv run memoryforge --db .memoryforge\memory.db lcm-context `
--session-id <session-id-from-lcm-sessions>
If lcm-sessions shows message_count: 0, the hook was not trusted, Codex was
not restarted after installing hooks, or Codex did not run from the initialized
project directory.
- Optional semantic vector recall. Leave this off for the first smoke test. To enable FastEmbed later, set the environment before indexing and before starting Codex:
$env:MEMORYFORGE_VECTOR_BACKEND='fastembed'
$env:MEMORYFORGE_VECTOR_MODEL='BAAI/bge-small-en-v1.5'
Without those variables, 6.1.5 uses lexical BM25/FTS recall by default so
first run stays responsive on Windows and offline machines.
Optional Codex MCP Adapter
For Codex CLI recall/context tools, register the MemoryForge MCP server:
codex mcp add memoryforge -- uv run memoryforge-mcp
Then run MemoryForge init at the project root:
uv run memoryforge init . --agent-id codex --force
This creates:
.memoryforge/memory.db
.memoryforge/config.json
AGENTS.md
During init, MemoryForge creates or updates the root AGENTS.md with a guarded
MemoryForge instruction block:
<!-- MemoryForge instructions start -->
...
<!-- MemoryForge instructions end -->
MemoryForge does not create project-local .codex/ files or install Codex
hooks unless --install-hooks is requested. It also does not call Codex CLI
subprocesses during init or project indexing. The MCP adapter intentionally
exposes only the hot-path tools:
recall_memory: fast factual recall from durable RLM/LTM indexesbuild_context_bundle: grounded LCM/LTM context assembly for the active modelindex_analyze: index Markdown and return host-subagent RLM analysis plans
LCM completed-turn capture is not exposed through MCP. For Codex CLI, install and trust the WSL/Linux hook runner so lifecycle capture is continuous instead of relying on ad-hoc tool calls.
index_analyze mirrors the internal RLM planner: it chunks/indexes selected
Markdown files, writes raw rlm_chunk:<id> evidence into LTM, and returns
plans[].batches[].host_subagent_prompt for the active Codex host to dispatch.
It does not call a model or spawn codex exec.
WSL/Linux Hook Auto-Capture
Use hooks when you need Codex interactive mode to auto-capture prompts, tool outputs, stop events, and compaction snapshots into LCM. For reliability, run Codex from WSL/Linux and install hooks there:
uv run memoryforge init . --agent-id codex --force --install-hooks
This creates .codex/hooks.json plus a tiny local runner:
.memoryforge/hooks/memoryforge-hook.sh
.memoryforge/hooks/memoryforge-hook.log
After restarting Codex from the same WSL/Linux project directory, run /hooks
and trust the MemoryForge hook definitions. Without that trust step, Codex will
skip project-local command hooks.
The hook is intentionally local-only:
- It calls
python -m memoryforge.cli.main hook .... - It does not call
codex. - It does not call a model.
- It does not run
codex exec. - It does not depend on the active Codex account or OAuth token.
- It exits
0and writes diagnostics to.memoryforge/hooks/memoryforge-hook.log.
LCM lifecycle capture is additive. RLM/LTM ingestion and recall keep working the same way; hooks only append active-session turns and context snapshots into the LCM tables.
Basic Usage
Default Codex CLI usage is WSL/Linux hook lifecycle capture plus MCP recall, context, and RLM planning. The CLI commands below are the direct/manual surface for indexing, recall, context inspection, and maintenance.
Ingest a long Markdown file or project document:
uv run memoryforge --db .memoryforge/memory.db ingest-file docs/notes.md \
--agent-id codex
Index project Markdown quickly:
uv run memoryforge index . \
--agent-id codex
Run RLM analysis only for large Markdown files from inside Codex through MCP:
Use MemoryForge MCP index_analyze for this project with analyze_min_bytes=20000 and analyze_max_files=5, then dispatch the returned host_subagent_prompt batches.
MCP index_analyze does not spawn codex exec or any external model process.
It returns plans[] with:
host_subagent_prompt: prompt for the active Codex host to give to a subagentfetch_command_argvs: exact chunk fetch commandsrecord_command_argv: exact command to record that batch analysisaggregate_command_argv: final aggregation command with--expected-batches
The intended flow inside Codex is:
- Call MCP
memoryforge.index_analyze. - For each returned batch, spawn a Codex host subagent using
host_subagent_prompt. - Each subagent fetches only its chunks and writes a concise cited analysis.
- Record each analysis with the returned
record_command_argv. - Run
aggregate_command_argvafter all planned batches are recorded.
For targeted/debug usage, rlm-run still exists as a legacy advanced CLI
command, but it is not the primary indexing path and is intentionally omitted
from the quickstart. Prefer MCP index_analyze plus host subagents.
The primary index path chunks sources losslessly and stores exact
rlm_chunk:<id> refs for deep rehydration. MCP index_analyze prepares
host-subagent analysis work and stores rlm_analysis/rlm_summary rows in LTM
after the returned record/aggregate commands are run.
Recall durable evidence:
uv run memoryforge --db .memoryforge/memory.db recall-memory \
--agent-id codex \
--query "why did we choose sqlite"
Build a runtime context bundle for the active Codex project:
uv run memoryforge --db .memoryforge/memory.db runtime-context \
--agent-id codex \
--session-id session-1 \
--query "what context should Codex use now" \
--project-root .
Run LCM compaction over MemoryForge's stored active context:
uv run memoryforge --db .memoryforge/memory.db lcm-compact \
--agent-id codex \
--session-id session-1 \
--project-root . \
--force
Inspect the active LCM state before or after compaction:
uv run memoryforge --db .memoryforge/memory.db lcm-sessions \
--agent-id codex
uv run memoryforge --db .memoryforge/memory.db lcm-messages \
--session-id session-1 \
--agent-id codex \
--include-content
uv run memoryforge --db .memoryforge/memory.db lcm-summary \
--session-id session-1
Run the MCP server directly:
uv run memoryforge-mcp
Vector Recall
MemoryForge uses lexical BM25/FTS recall by default. To enable semantic vector
recall with FastEmbed and store local embeddings in vec_index, configure:
export MEMORYFORGE_VECTOR_BACKEND=fastembed
export MEMORYFORGE_VECTOR_MODEL=BAAI/bge-small-en-v1.5
For explicit lexical-only fallback, set MEMORYFORGE_VECTOR_BACKEND=disabled.
The project intentionally keeps one vector cache table, vec_index, and avoids
SQLite extension backends such as sqlite-vec in the core release. This keeps
the package easier to install, test, and publish.
Retrieval is hybrid by design: vector recall and lexical recall can both contribute candidates, and MemoryForge fuses bounded evidence for the runtime context instead of relying on a vector-only path.
CLI Surface
Public commands:
- Project/runtime:
init,mcp-server,runtime-context - Conversation memory:
store-session,search,recall-memory,active-recall,long-term-source - Contradictions:
record-contradiction,find-contradictions - LCM:
lcm-context,lcm-sessions,lcm-messages,lcm-summary,lcm-compact,lcm-maintain - RLM/source loading:
ingest-file,rlm-load,rlm-search,rlm-chunk-get,dispatch,context-get,rlm-record,aggregate,rlm-run - Diagnostics:
chunk,benchmark
MCP intentionally exposes only recall_memory, build_context_bundle, and
index_analyze. Low-level RLM/debug commands stay on the CLI so their schemas
are not injected into every Codex turn.
memoryforge hook remains available as an internal endpoint for direct testing.
RLM host-subagent analysis is coordinated by MCP index_analyze: the active
Codex host owns subagent execution, while MemoryForge owns chunk storage,
record, aggregate, and recall.
Benchmarks
The current benchmark focus is long-memory behavior, not static code indexing:
- LoCoMo
- LongMemEval
- deterministic multi-session stress benchmark
- synthetic smoke benchmark
Example smoke check:
uv run python benchmarks/synthetic_test.py
Stress check for many real SQLite sessions:
uv run python benchmarks/stress_sessions.py \
--sessions 100 \
--turns-per-session 12 \
--output benchmarks/results/stress_sessions_100x12.json
Real LoCoMo and LongMemEval runs require their datasets and model credentials. See docs/BENCHMARKS.md for run modes and result fields.
Development
Run the normal quality gate on Unix-like shells:
make check
Equivalent commands:
uv run ruff check memoryforge tests benchmarks
PYTHONDONTWRITEBYTECODE=1 uv run mypy memoryforge
PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=. uv run pytest --ignore=tests/test_real_subagents.py --cov=memoryforge --cov-report=term-missing --cov-fail-under=77
MEMORYFORGE_REAL_SUBAGENT=1 MEMORYFORGE_REAL_PROJECT_ROOT="$PWD" MEMORYFORGE_SUBAGENT_RUNNER=codex MEMORYFORGE_MODEL=gpt-5.4 uv run pytest tests/test_real_subagents.py -vv
uv build
uv run twine check dist/*
On Windows PowerShell, keep pytest temp/cache paths in writable directories:
$env:TMP='C:\tmp'; $env:TEMP='C:\tmp'
Remove-Item Env:\MEMORYFORGE_SUBAGENT_RUNNER -ErrorAction SilentlyContinue
Remove-Item Env:\MEMORYFORGE_MODEL -ErrorAction SilentlyContinue
uv run pytest --ignore=tests/test_real_subagents.py --basetemp=C:\tmp\memoryforge-pytest-basetemp -o cache_dir=.tmp\pytest-cache
$env:MEMORYFORGE_REAL_SUBAGENT='1'; $env:MEMORYFORGE_REAL_PROJECT_ROOT=(Get-Location).Path
$env:MEMORYFORGE_SUBAGENT_RUNNER='codex'; $env:MEMORYFORGE_MODEL='gpt-5.4'
uv run pytest tests/test_real_subagents.py -vv --basetemp=C:\tmp\memoryforge-pytest-basetemp-real -o cache_dir=.tmp\pytest-cache-real
The real Codex sub-agent smoke tests are local-only. Run them on a machine with the Codex CLI installed and authenticated if you want to verify runner="codex"; they are not part of CI/CD. Mock runners are only for targeted unit tests that verify MemoryForge's own control flow.
Release Notes For Maintainers
Before pushing or publishing:
- Keep generated data out of the release:
.venv/, caches,.coverage,.memoryforge/,.codebase-memory/,dist/, andbenchmarks/results/. - Run the full gate on Python 3.10, 3.11, and 3.12 through CI.
- Build the wheel and sdist with
uv build. - Check distributions with
twine check. - Prefer the
PublishGitHub workflow with PyPI trusted publishing. Direct maintainer uploads may usetwine uploadwithTWINE_USERNAME=__token__andTWINE_PASSWORDsupplied from the shell environment, never from a committed config file.
Documentation
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file memfg-6.1.5.tar.gz.
File metadata
- Download URL: memfg-6.1.5.tar.gz
- Upload date:
- Size: 196.1 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.12.13
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
0c0ae6e0a28c90249dfe638dfe2ee688b597d6d7e8454438125bb923461d03ae
|
|
| MD5 |
aeb57fe96da98e67ae51feca3e72ccc6
|
|
| BLAKE2b-256 |
70630af7e08e73f9b07fb30040d8a4bd25a06e9c992fcede5c22e1db07e1fc2f
|
File details
Details for the file memfg-6.1.5-py3-none-any.whl.
File metadata
- Download URL: memfg-6.1.5-py3-none-any.whl
- Upload date:
- Size: 164.3 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.12.13
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
70c596f0752c4e8d054b5871080c6a437b2b6cc57334584aee19eaa9d00d65ef
|
|
| MD5 |
a7685fb3c87ccebcfdb2ed6a5a5c83a7
|
|
| BLAKE2b-256 |
204f6e6da67c689a8546f681c4af1bf3f37cca259ddfbbecf80f4c5cac3c6ffc
|