lossless-claude
Shared memory infrastructure for coding agents

DAG-based summarization, SQLite-backed message persistence, promoted long-term memory, MCP retrieval tools

Website • Runtime Model • Installation • MCP Tools • Development

---

lossless-claude replaces sliding-window forgetfulness with a persistent memory runtime for both humans and agents.

• Every message is stored in a project SQLite database.
• Older context is compacted into a DAG of summaries instead of being dropped.
• Durable decisions and findings are promoted into cross-session memory.
• Claude Code already has end-to-end hook integration, while VS Code and Codex use connector-based workflows on the same backend today.

Humans and agents use the same backend. The integration surface differs by client, but the memory model is shared.

This repo started as a fork of lossless-claw by Martian Engineering, adapted for Claude Code. The LCM model and DAG architecture originate from the Voltropy paper.

Runtime Model

flowchart LR
  subgraph Clients["Clients"]
    CC["Claude Code<br/>hooks + MCP"]
  end

  CC --> D["lossless-claude daemon"]

  D --> DB[("project SQLite DAG")]
  D --> PM[("promoted memory FTS5")]
  D --> TOOLS["MCP tools<br/>search / grep / expand / describe / store / stats / doctor"]

Loading

Capabilities by integration path

Path	Restore	Prompt hints	Turn writeback	Automatic compaction	Notes
Claude Code	Yes	Yes	Yes, via transcript/hooks	Yes	Primary hook-based integration
GitHub Copilot (VS Code)	No	Yes, via skill/rules	No	No	Repo-local skill can teach Copilot to call lcm, but there is no automatic restore or turn capture yet
Codex	No	Yes, via skill/rules	No	No	Repo-local or global skill plus lcm import --codex; MCP config in .codex/config.toml is still manual, and first-class runtime support is tracked in issue #232

LCM Model

Phase	What happens
Persist	Raw messages are stored in SQLite per conversation
Summarize	Older messages are grouped into leaf summaries
Condense	Summaries roll up into higher-level DAG nodes
Promote	Durable insights are copied into cross-session memory
Restore	New sessions recover context from summaries and promoted memory
Recall	Agents query, expand, and inspect memory on demand

Nothing is dropped. Raw messages remain in the database. Summaries point back to their sources. Promoted memory remains searchable across sessions.

flowchart TD
  A["conversation / tool output"] --> B["persist raw messages"]
  B --> C["compact into leaf summaries"]
  C --> D["condense into deeper DAG nodes"]
  C --> E["promote durable insights"]
  D --> F["restore future context"]
  E --> F
  F --> G["search / grep / describe / expand / store"]

Loading

Installation

Prerequisites

• Node.js 22+
• Claude Code if you want hook-based automation
• GitHub Copilot in VS Code if you want VS Code integration
• Codex CLI if you want Codex connector installation, summarization, or transcript import

Claude Code

Install the lcm binary first:

npm install -g @lossless-claude/lcm  # provides the `lcm` command

claude plugin add github:lossless-claude/lcm
lcm install

lcm install writes config, registers hooks, installs slash commands, registers MCP, and verifies the daemon.

VS Code (GitHub Copilot)

Install the lcm binary first:

npm install -g @lossless-claude/lcm

Then install the repo-local Copilot connector:

lcm connectors install github-copilot
lcm connectors doctor github-copilot

This creates a workspace skill under .github/skills/lcm-memory/SKILL.md so Copilot can search and store memory through the lcm CLI.

Codex

Install the lcm binary first:

npm install -g @lossless-claude/lcm

Then install the Codex connector:

lcm connectors install codex
lcm connectors doctor codex

Import older Codex sessions when needed:

lcm import --codex

If you also want MCP inside Codex, run lcm connectors install codex --type mcp. Today that prints the TOML block you must add manually to .codex/config.toml.

See docs/vscode-codex.md for the current VS Code/Codex setup path and known shortcomings.

Hooks

Claude Code uses four hooks. All hooks auto-heal: each validates that all required entries remain registered and repairs missing entries before continuing.

Hook	Command	Purpose
PreCompact	lcm compact --hook	Intercepts compaction and writes DAG summaries
SessionStart	lcm restore	Restores project context, recent summaries, and promoted memory
SessionEnd	lcm session-end	Ingests the completed Claude transcript
UserPromptSubmit	lcm user-prompt	Searches memory and injects prompt-time hints

flowchart LR
  SS["SessionStart"] --> CONV["Conversation"]
  CONV --> UP["UserPromptSubmit<br/>(each prompt)"]
  UP --> CONV
  CONV --> PC["PreCompact<br/>(if context fills)"]
  PC --> CONV
  CONV --> SE["SessionEnd"]

Loading

MCP Tools

Tool	Purpose
lcm_search	Hybrid search across episodic memory (SQLite) and semantic memory
lcm_grep	Regex or full-text search across raw messages and summaries
lcm_expand	Decompress a summary node into its source content by traversing the DAG
lcm_describe	Inspect metadata and lineage of a memory node (depth, token count, parent/child links)
lcm_store	Persist durable memory manually with optional tags
lcm_stats	Show token savings, compression ratios, and usage statistics
lcm_doctor	Diagnose daemon, hooks, MCP registration, and summarizer setup

CLI

# Setup & diagnostics
lcm install                # setup wizard
lcm uninstall              # remove hooks, MCP, and config
lcm doctor                 # diagnostics: daemon, hooks, MCP, summarizer
lcm diagnose               # scan recent sessions for hook failures
lcm status                 # daemon + summarizer mode
lcm -V                     # version

# Memory inspection
lcm search "query"        # search episodic and promoted memory
lcm grep "pattern"        # search messages and summaries
lcm describe <nodeId>      # inspect metadata for a memory node
lcm expand <nodeId>        # expand a summary node into source detail
lcm store "content"       # persist a durable memory entry
lcm stats                  # memory and compression overview
lcm stats -v               # per-conversation breakdown
lcm stats --pool           # connection pool statistics

# Compaction & promotion
lcm compact                # compact the current project
lcm compact --all          # compact all tracked projects
lcm promote                # promote durable insights to long-term memory
lcm promote --all          # promote across all tracked projects

# Import / export
lcm import                 # import Claude Code sessions for the current project
lcm import --all           # import all projects
lcm export                 # export promoted knowledge to JSON
lcm import-knowledge <f>   # import a knowledge JSON file

# Connectors (wire lcm into other AI agents)
lcm connectors list        # list available agents and installed connectors
lcm connectors install <a> # install a connector for an agent
lcm connectors remove <a>  # remove a connector for an agent
lcm connectors doctor      # check connector health

# Sensitive data
lcm sensitive add <pat>    # add a redaction pattern (project-scoped)
lcm sensitive add --global # add a global redaction pattern
lcm sensitive list         # list all active patterns
lcm sensitive test <str>   # test what gets redacted
lcm sensitive purge --yes  # remove all stored data for the current project

# Daemon
lcm daemon start --detach  # start daemon in background

# Hook handlers (internal — called by Claude Code hooks)
lcm compact --hook         # PreCompact hook
lcm restore                # SessionStart hook
lcm session-end            # SessionEnd hook
lcm user-prompt            # UserPromptSubmit hook
lcm post-tool              # PostToolUse hook (passive learning)

# MCP server
lcm mcp                    # start MCP server

Configuration

All environment variables are optional. The default summarizer mode is auto.

Variable	Default	Description
LCM_SUMMARY_PROVIDER	auto	auto, claude-process, codex-process, anthropic, openai, or disabled
LCM_SUMMARY_MODEL	unset	Optional model override for the selected summarizer provider
LCM_CONTEXT_THRESHOLD	0.75	Context fill ratio that triggers compaction
LCM_FRESH_TAIL_COUNT	32	Most recent raw messages protected from compaction
LCM_LEAF_MIN_FANOUT	8	Minimum raw messages per leaf summary
LCM_CONDENSED_MIN_FANOUT	4	Minimum summaries per condensed node
LCM_INCREMENTAL_MAX_DEPTH	0	Automatic condensation depth
LCM_LEAF_CHUNK_TOKENS	20000	Maximum source tokens per leaf compaction pass
LCM_LEAF_TARGET_TOKENS	1200	Target size for leaf summaries
LCM_CONDENSED_TARGET_TOKENS	2000	Target size for condensed summaries
LCM_MAX_EXPAND_TOKENS	4000	Token cap for DAG expansion via lcm_expand
LCM_LARGE_FILE_TOKEN_THRESHOLD	25000	File size (tokens) above which content is extracted to disk
LCM_AUTOCOMPACT_DISABLED	false	Set to true to disable automatic compaction after each turn
LCM_ENABLED	true	Set to false to disable the plugin while keeping it registered

auto resolves per caller:

• lcm -> claude-process
• explicit config or LCM_SUMMARY_PROVIDER override always takes precedence

See docs/configuration.md for tuning notes and deeper operational guidance.

Development

npm install
npm run build
npx vitest
npx tsc --noEmit

Repository layout

bin/
  lcm.ts                      CLI entry point (binary: lcm)
src/
  compaction.ts               DAG compaction engine
  connectors/                 client integration adapters
  daemon/                     HTTP daemon, lifecycle, config, routes
  db/                         SQLite schema + promoted memory
  hooks/                      Claude hook handlers + auto-heal
  llm/                        summarizer backends
  mcp/                        MCP server + tool definitions
  store/                      conversation and summary persistence
installer/
  install.ts                  setup wizard
  uninstall.ts                cleanup
test/
  ...                         Vitest suites

Privacy

All conversation data is stored locally in ~/.lossless-claude/. Nothing is sent to any lossless-claude server.

If you configure an external summarizer (claude-process, anthropic, openai, etc.), messages are sent to that provider for summarization — after built-in secret redaction. lossless-claude scrubs common secret patterns (API keys, tokens, passwords) from message content before writing to SQLite and before sending to the summarizer.

Add project-specific patterns with lcm sensitive add "MY_PATTERN". See docs/privacy.md for full details.

Technical Notes

• Claude Code integration is hook-first.
• The daemon is shared; the memory backend is client-agnostic.
• The repo carries the original lossless-claw lineage; the current runtime is Claude Code oriented.

Acknowledgments

lossless-claude stands on the shoulders of lossless-claw, the original implementation by Martian Engineering. The DAG-based compaction architecture, the LCM memory model, and the foundational design decisions all originate there.

The underlying theory comes from the LCM paper by Voltropy.

License

MIT