claude-lcm

claude-lcm

Lossless Context Management for Claude Code — a native skill port of the LCM paper (Voltropy, Feb 2026).

Claude Code natively truncates older messages when the context window fills up. claude-lcm replaces that with a deterministic, lossless memory architecture: every message is persisted verbatim in SQLite, older messages are compacted into a hierarchical DAG of summaries, and the agent can always retrieve any prior state.

No OpenClaw required. Works natively in Claude Code as a skill.

---

What it does

When a session grows beyond comfortable context size, claude-lcm:

1. Persists every message verbatim in SQLite — the immutable store
2. Compacts older messages into leaf summary nodes using depth-aware LLM prompts
3. Condenses summaries into higher-level DAG nodes as they accumulate
4. Assembles active context each turn: summaries + recent raw messages (fresh tail)
5. Provides retrieval tools (lcm_grep, lcm_describe, lcm_expand_query) so the agent can search and recall details from compacted history
6. Processes large datasets via lcm_llm_map and lcm_agentic_map (operator-level recursion)
7. Writes checkpoints for cross-session continuity

Nothing is lost. The agent that never forgets — because it doesn't.

---

Architecture

Inspired by the LCM paper and the lossless-claw OpenClaw plugin.

Raw messages → Leaf summaries (d0) → Condensed (d1) → Condensed (d2) → ...
                     ↑                      ↑
              summary_messages        summary_parents

Dual-state memory:

• Immutable Store — SQLite DB, verbatim message history, never modified
• Active Context — assembled from summary nodes + protected fresh tail

Three-level summarization escalation (guaranteed convergence):

1. Detailed narrative (LLM)
2. Bullet points, half target tokens (LLM)
3. Deterministic truncation to 512 tokens (no LLM, never fails)

See .claude/skills/claude-lcm/references/architecture.md for the full schema and algorithms.

---

Installation

Via Claude Code Plugin Marketplace (recommended)

# In Claude Code, add the marketplace
/plugin marketplace add srikalyan/srikalyan-marketplace

# Install the plugin
/plugin install claude-lcm@srikalyan-marketplace

Manual installation

# Clone and run the install script
git clone https://github.com/srikalyan/claude-lcm.git
cd claude-lcm
./install.sh

Prerequisites

• Claude Code (with claude CLI on PATH)
• python3 (stdlib only, no pip installs needed)

---

Usage

Initialize a session

python3 scripts/lcm_init.py
source ~/.claude-lcm/session.env

Check context health

python3 scripts/lcm_status.py

Compact when context is filling up

python3 scripts/lcm_compact.py          # full pass (leaf + condensation)
python3 scripts/lcm_compact.py --mode leaf  # leaf only
python3 scripts/lcm_compact.py --dry-run   # preview without writing

Search history

python3 scripts/lcm_grep.py "database migration"
python3 scripts/lcm_grep.py "API threshold" --scope summaries
python3 scripts/lcm_grep.py "config" --mode regex --limit 20

Inspect a summary or file

python3 scripts/lcm_describe.py sum_abc123
python3 scripts/lcm_describe.py file_xyz789

Answer a focused recall question

python3 scripts/lcm_expand_query.py \
  --prompt "What was the final decision on the compaction threshold?" \
  --query "compaction threshold"

Process a dataset (LLM-Map)

python3 scripts/lcm_llm_map.py \
  --input data.jsonl \
  --output results.jsonl \
  --prompt "Extract entities from this text" \
  --schema schema.json \
  --concurrency 16

Process with multi-step reasoning (Agentic-Map)

python3 scripts/lcm_agentic_map.py \
  --input tasks.jsonl \
  --output results.jsonl \
  --prompt "Analyze this repository" \
  --concurrency 4 --read-only

Session handoff

# Before ending a long session
python3 scripts/lcm_checkpoint.py

# On next session start
python3 scripts/lcm_resume.py

---

Configuration

Set via environment variables before running scripts (or add to your shell profile):