Memory bank

The Memory Bank is where Forven keeps what it has learned. It is a searchable store of quant skills — validated trading insights distilled from your backtests — plus the unconfirmed observations still gathering evidence. The store grows as strategies run, and what it holds steers future research.

This page is for anyone using the Memory Bank to explore, search, and tidy that knowledge. It covers the /memory page: the tier system, search and filters, the consolidation pipeline, and outcome tracking. For the mechanics behind how skills are born, scored, and fed back into ideation, see quant skills.

Forven is a research tool. The Memory Bank records patterns observed in historical and simulated data; none of it predicts the future, and nothing here is financial advice.

What lives in the Memory Bank

The store holds two kinds of thing, at two levels of certainty:

• Quant skills — validated, versioned insights. Each carries a skill_type (regime, failure, indicator, combo, params), a confidence score from 0 to 1, an evidence list, and what_works / what_doesnt_work notes. Skills are the durable layer.
• Candidate hypotheses — unconfirmed observations extracted from backtest results. An observation needs three backtests of supporting evidence before it promotes to a skill. These are the provisional layer, on probation.

Both are read by Forven's ideation step, so the store is not a passive log — it is working knowledge that shapes what strategies get proposed next.

The word hypothesis is overloaded. The candidate hypotheses here are quant-skill observations on the way to becoming skills. They are not the same as the research-record hypotheses (crucibles) in the Hypotheses Manager, even though both use the word. Context disambiguates; this page is about the learning store.

The tier system

Skills sort into tiers by how trusted they are. The Memory Bank lets you filter by tier so you can separate raw noise from settled knowledge:

• Untiered — newly extracted, not yet promoted.
• Signal — an early observation; one or two data points.
• Working — corroborated across enough evidence to influence ideation.
• Canon — the most trusted, highest-confidence knowledge.

Tier rises with accumulated evidence and confidence; it is not a popularity ranking. A skill that keeps being right in out-of-sample tests climbs; a skill that fails in live use loses confidence and can fall.

Exploring the store

The Memory Bank opens on the Explore view. Use it to find what Forven already knows about a pattern before you invent a strategy from scratch.

Filters available in Explore:

• Search — free text over skill names and bodies (for example mean reversion or funding).
• Source — Workspace, Chroma (the vector memory), or Narratives.
• Tier — Untiered, Signal, Working, Canon.
• Time range — All, 24H, 7D, 30D, 90D.
• Agent / strategy — narrow to skills tied to a specific agent or strategy.

Selecting an item opens the detail drawer, which shows the skill's timeline, annotations, and a confidence sparkline so you can see how trust in that skill has moved over time.

Steps: search and read a skill

1. Open the Memory Bank at /memory.
2. Type a query in the search box — for example mean reversion.
3. Narrow with the source checkboxes (Workspace, Chroma, Narratives) and the agent/strategy dropdowns.
4. Pick a time range preset (All, 24H, 7D, 30D, 90D).
5. Filter by tier to separate Signal noise from Canon knowledge.
6. Click a result to open the detail drawer with its timeline and confidence sparkline.

What you'll see: a filtered list on the left and, on selection, a right-hand drawer with the skill's full body, evidence, version history, and a sparkline of its confidence over time.

The candidate pipeline and consolidation

The Skill Candidates tab (the consolidation pipeline) is where provisional observations queue up. It shows the candidate hypotheses, pipeline stats, and a Consolidate action.

Consolidation merges similar skills by similarity so the store stays tight instead of sprawling into near-duplicates. This is also where periodic housekeeping retires low-value knowledge: skills that have stayed below a confidence floor across enough samples are archived, and stale candidate observations are pruned after their evidence window lapses.

Steps: consolidate skill candidates

1. From /memory, open the Skill Candidates tab.
2. Review the pending candidates and pipeline stats.
3. Click Consolidate to merge similar skills and clean the store.

What you'll see: the candidate list, consolidation statistics, and the consolidate button. After consolidation, near-duplicate skills are merged into a single, better-evidenced entry.

Consolidation merges by similarity and is not reversible. Preview the candidates before you run it — there is no undo once skills are merged.

Outcome tracking

A skill is only as good as its track record, so the Memory Bank closes the loop with real outcomes. When a strategy reaches a terminal state — archived, retired, or live_graduated — Forven looks up which skills that strategy cited and records the result against them.

A skill's outcomes view shows recent verdicts and a win/loss history. Each outcome nudges the skill's confidence: a positive result lifts it, a negative result lowers it more sharply, and a neutral result leaves it flat. The exact deltas and decay rules — confidence weighted toward recent evidence, the consolidation thresholds, the promotion count — are documented in quant skills.

This is the discipline that keeps the store honest. A pattern that looked good in one backtest but failed repeatedly in paper and live trading will quietly lose its standing, regardless of how confident it once seemed.

Confidence movement is illustrative of a skill's track record in Forven's own tests, not a forecast. A high-confidence skill is one that has held up so far — not a promise that it will keep working.

Maintenance: forget and prune

The Memory Bank includes maintenance controls for pruning the store directly: a forget action for individual items and bulk-prune for clearing out stale or low-value knowledge in batches. Use these when the store accumulates noise you do not want feeding back into ideation. As with consolidation, treat removal as permanent.

Memory Bank vs. the Brain's memory

The Memory Bank is the long-term knowledge store for quant skills. It is not the same as the working memory the brain keeps for itself:

• Memory Bank (/memory) — durable, searchable quant skills built up over many strategies. Shared knowledge.
• Brain working notes (/brain) — the brain's own short, operational notes carried between decision cycles, plus its self-judged lessons. Private to the orchestrator.

If you are looking for the patterns Forven has learned about markets, you want the Memory Bank. If you are looking at how the orchestrator is reasoning right now, you want the Brain page.

Caveats

• The candidate pipeline is provisional by design. A Signal-tier observation is a hint, not a conclusion — trust rises only as out-of-sample evidence accumulates.
• Consolidation, forget, and bulk-prune are not reversible. Preview before you run them.
• Confidence scores reflect performance in Forven's backtests and simulations. They describe the past; they do not predict it.

Related

• Quant skills (the learning loop) — how skills are extracted, scored, and fed back into ideation.
• The brain — the orchestrator's own working memory and self-judged lessons.
• Hypothesis-driven research — the research-record hypotheses (crucibles), distinct from candidate skills.