Quick definition
Treeova's three-layer agent memory system: append-only message ledger, RL-aware recursive summarization, and hybrid full-text + semantic retrieval. Lets long-running agents retain decision-grade signal across sessions without exceeding model context windows.
The three layers
Layer one is an append-only ledger — every tool call, every observation, every decision is stored verbatim with a stable id. Layer two runs recursive summarization that gives extra weight to events the RL feedback loop marked as decision-grade. Layer three is hybrid retrieval: trigram + full-text search for exact recall plus semantic embeddings for related context.
Why "lossless" matters for trading agents
An options agent that forgets it already trimmed a position at +20% will happily trim again, blowing through size limits. By guaranteeing that summarization never deletes the underlying ledger row, LCM lets agents safely run for weeks while keeping each individual LLM call inside a tight token budget.
Step-aware thinning
Inside a single multi-step reasoning chain, LCM applies step-aware thinning starting at step 8 — earlier steps get progressively summarized so the agent doesn't suffer from late-step token starvation. The thinning is reversible because the original ledger is still on disk.
Go deeper
The full technical methodology behind Lossless Context Management (LCM) is documented in the Treeova whitepaper series.
Read the Lossless Context Management (LCM) whitepaper →