WP-03 — Lossless Context Management (LCM): Infinite Agent Memory

Long-running trading agents face a fundamental constraint: model context windows are finite, but the history that matters to a good decision is not. Lossless Context Management (LCM) is Treeova's answer — a closed-loop memory system that lets an agent reason over months of activity while still fitting inside a single model call.

The "lossless" claim is structural: raw inputs are never deleted. Compression happens only in derived summaries, which can always be rebuilt from the underlying ledger when the agent needs nuance the current summary lacks.

See the product surface for this methodology: Feature page for Lossless Context.

LCM is organized as three cooperating layers:

• Raw ledger. An append-only record of every agent-relevant message — user input, agent reasoning, tool calls, tool results. Nothing in this layer is ever overwritten or deleted.
• Recursive summaries. A tree of derived summaries that compress contiguous slices of the ledger into successively higher-level views. Summaries are themselves summarized when they grow, producing a hierarchy that scales with history length.
• Assembled context. The working window the model actually sees on a given call — composed of recent raw messages, relevant summaries pulled by retrieval, and any pinned system material the agent depends on.

The raw ledger is the source of truth. When older messages would exceed the assembled context, they are summarized into the recursive layer — they are never removed from the ledger. This is the basis of the "lossless" property: any summary's nuance can be recovered, because the underlying messages still exist.

Because the ledger is append-only, audit and replay are first-class operations. The exact sequence of inputs that produced a past decision can always be reconstructed.

As history grows, contiguous slices of the ledger are compressed into summaries. As summaries grow, they are themselves compressed. The result is a hierarchy in which a single high-level summary can cover a long stretch of history while finer-grained children remain available for selective retrieval.

Summaries are not snapshots — they are re-derivable. If a downstream decision reveals that an earlier summary missed material that turned out to matter, a fresh summary can be regenerated from the underlying ledger slice without disturbing the rest of the hierarchy.

A naive summarizer treats reinforcement-learning signals as background noise and smooths them out. LCM's compaction is structured: each summary carries explicit RL metadata — outcome tags, calibration signals, lifecycle markers — alongside its prose body.

When summaries are summarized again, those structured signals propagate upward rather than being averaged away. The calibration loop downstream (see WP-10) can therefore consume summary-derived signal without losing the phase-aware fidelity it depends on.

When an agent needs old context, LCM retrieves it through two complementary channels in parallel: full-text search over the raw ledger, and semantic similarity over summary embeddings. Full-text finds exact matches the agent literally said or saw; semantic finds passages whose meaning resembles the current question even if the wording differs.

Both result sets are merged into the assembled context with provenance — the agent can tell whether a snippet is raw or summarized, and the summary level it came from.

• The specific depth thresholds at which compaction is triggered.
• The model routing used to produce summaries at each depth.
• Summary token budgets per layer.
• The exact prompts used by the summarization operators.
• Escalation policies for compaction failures.

• LCM is a memory architecture, not a reasoning upgrade. It cannot make a base model smarter; it can only ensure the model has the right history in front of it.
• Retrieval quality depends on how distinctively past events were recorded. Vague or repetitive ledger entries are harder to find later, regardless of the retrieval channel.
• A summary derived once can lose nuance; LCM's safeguard is re-derivation from the ledger, not infallible first-pass compression.
• The system assumes the underlying message stream is itself faithful. LCM cannot correct hallucinations or omissions that were never written into the ledger to begin with.