LCM Prototype

A Python prototype implementing core ideas from Lossless Context Management (LCM) for long-horizon agent memory and context control.

Original Paper

📄 LCM: Lossless Context Management

• Author: Clint Ehrlich, Voltropy PBC
• Date: 2026-02-14
• PDF: docs/LCM-paper-voltropy-2026.pdf
• Source: https://papers.voltropy.com/LCM

Key Ideas

LCM emphasizes preserving original information while enabling bounded-context operation via:

• Immutable raw record storage
• Hierarchical summaries with pointers back to originals
• Multi-stage compaction with deterministic fallback
• Active-context budgeting with soft/hard thresholds

Architecture Overview

• ImmutableStore (store.py): append-only JSONL message log with query APIs.
• SummaryDAG (dag.py): hierarchical summary graph preserving pointers to source messages.
• Compactor (compactor.py): normal / aggressive / deterministic compaction levels.
• ContextManager (context.py): active window control via tau_soft, tau_hard.
• LCMEngine (engine.py): orchestrates ingestion, storage, context update, and retrieval.
• Supporting modules:
  • file_handler.py: large file registration + exploration summaries
  • operators.py: llm_map / agentic_map style operators
  • delegation.py: anti-infinite-delegation guardrails

Development

uv venv --python 3.14 .venv
uv pip install pytest tiktoken
pytest

Benchmarks

Custom Benchmark (100-round conversational recall)

Questions asked mid-conversation at Round 85 and 95. Qwen3.5+ as LLM Judge.

System	Round 85	Round 95	Overall
Atlas v2	0.75	0.75	0.75
OpenViking	0.75	0.75	0.75
LCM Prototype	0.74	0.74	0.74
Nowledge Mem	0.62	0.62	0.62
Atlas baseline	0.50	0.50	0.50

→ Detailed Report

LoCoMo Standard Benchmark (snap-research/locomo, ACL 2024)

Questions asked after full conversation (retrieval over stored history). 3 conversations, 252 QA pairs.

System	Overall	single-hop	temporal	multi-hop	open-domain	adversarial
Atlas baseline	0.434	0.390	0.317	0.357	0.817	0.233
Atlas v2	0.299	0.383	0.275	0.381	0.398	0.125
OpenViking	0.206	0.436	0.033	0.405	0.227	0.092
LCM	0.145	0.285	0.172	0.143	0.097	0.050
Nowledge	0.004	0.000	0.000	0.000	0.000	0.017

→ Detailed Report

Key Insight

Compression dominates during active conversation (custom benchmark), but raw semantic search wins for post-conversation retrieval (LoCoMo). The optimal memory system needs both modes.