Settings reference

How Always Remember reaches the model each turn. Full list puts every line in the system prompt (simple; cost grows with the list). Relevant only embeds your message and injects just the best-matching facts — scales to hundreds.

After the summarizer adds facts in the background, run a normalized + semantic near-duplicate merge automatically.

tune: Leave on unless you’re debugging fact writes.

Every N successful chat turns, a summarizer-model pass merges paraphrased facts (“wife is Sarah” + “married to Sarah” → one line). 0 disables.

Sampling randomness for the selected role: lower = more deterministic and factual, higher = more varied and creative. Disabled automatically for models that don’t support it.

tune: 0.2–0.4 suits summarizer / verifier roles; 0.6–0.8 suits main chat.

Cap on a single response’s length.

tune: Lower to force brevity and bound cost; raise if long answers get cut off mid-thought.

Blends keyword (BM25) and semantic (vector) matching into one score. Off = pure semantic.

tune: Turn off only to A/B pure-vector recall.

Keyword share of the blended score. 0.4 means 40% keyword, 60% semantic.

tune: Raise toward 0.6 for exact-term content (codes, part numbers, names); lower toward 0.2 for conversational content.

Minimum similarity before a retrieved memory is used at all.

tune: Raise if irrelevant snippets pollute answers; lower if Om-E “can’t find” content you know is there.

How many retrieved hits are injected per turn.

tune: Lower to save tokens; raise for broad synthesis questions.

Cross-encoder second pass over the top ~100 candidates — noticeably better precision, slower queries.

tune: Turn off on weak hardware if searches feel sluggish.

Include neighbouring chunks around every match so answers get surrounding context.

tune: Off if answers quote too much irrelevant surrounding text.

How many neighbour chunks each side when Parent Expansion is on.

tune: Raise for dense technical docs where one chunk rarely stands alone.

Content above this many characters gets chunked instead of stored whole.

tune: Lower to chunk more aggressively — finer retrieval, more vectors.

Upper size of each embedded chunk.

tune: Smaller chunks = more precise hits but less context per hit.

Context carried between neighbouring chunks so sentences aren’t cut dead at boundaries.

tune: Lower to shrink index size at a slight recall cost.

Per-chunk embeddings for large payloads (vs one embedding per document).

tune: Leave on; off is a debugging escape hatch.

Prefixes each chunk with [KB | Topic | Date] before embedding — better precision on large multi-topic KBs.

tune: Off only for single-topic KBs where headers add noise.

Embedding batch size — throughput vs memory.

tune: Lower on memory-constrained machines; raise on a strong GPU.

FAISS index. SQ8: quantized, 4× smaller, near-identical recall (recommended). Flat: exact fp32. HNSW: approximate, for stores over ~10K vectors. Applies on the next index rebuild.

tune: flat to rule out quantization in a recall investigation; hnsw for very large KBs.

Master switch. Off (or no graph built yet) = pure vector search everywhere.

tune: Off to A/B whether the graph is helping a given KB.

Chunks = vector only. Graph Enhanced = vector + 1-hop entity expansion + merged scoring.

tune: chunks if graph results feel tangential for a KB.

Auto-run entity extraction whenever content lands (files, pages, YouTube, topics). Off = the graph only grows manually.

tune: Off to speed up bulk ingests; extract later.

Score multiplier for results that arrive via the graph alone (no direct vector match). Lower = graph-only results rank lower.

tune: Lower if graph neighbours crowd out direct matches.

Minimum score lift for a vector hit that matches a graph entity.

tune: Raise slightly if entity-heavy questions under-rank.

Cap on the total graph lift — prevents runaway boosting.

tune: Lower if graph matches dominate too hard.

Expansion depth from matched nodes. 2+ hops only kick in when 1 hop finds too little.

tune: 2 for sparse graphs where relations chain (A → B → C).

Candidate pool fed into graph scoring — bigger pool, better coverage, slower query.

tune: Raise for very large KBs where relevant chunks miss the top-100.

Include other chats’ history (the cross-chat layer) in retrieval, subject to project-membership rules.

tune: Off for strict chat isolation.

How many recent tool actions (open tab, list tabs…) stay in the prompt so the model knows what it just did.

tune: Raise for long tool-heavy sessions; lower to save tokens.

Chat messages before a rolling summary fires in the background.

tune: Raise for short chats; lower to compress aggressively.

If token usage hits this first, the summary runs early — before the message-count trigger.

tune: Lower if long individual messages bloat context before the count trips.

Summaries kept in the prompt before older ones consolidate into one.

tune: Raise to preserve more long-range chat memory verbatim.

Loaded-topic message count that triggers batch mode.

tune: Lower if big loaded topics blow the context window.

Token estimate that triggers batch mode — whichever threshold trips first wins.

tune: Raise for million-token-window models; lower for small local models.

Target size of each chapter chunk.

tune: Bigger chapters = fewer turns, heavier each.

Preferred floor before a chunk is split (final tail excepted). Keep below Target.

Ceiling before a forced split. Keep above Target.

Auto-size chapters from the active model’s real context window instead of the fixed targets.

tune: Leave on — the fixed targets are the fallback.

Fraction of the model’s effective context used per chapter when adaptive sizing is on.

tune: Raise for huge-window models to finish in fewer turns.

Context held back for everything that isn’t the loaded payload — system prompt, history, tools.

tune: Raise if batch turns overflow; lower for lean chats.

Chapters processed on each turn of a running large-topic job.

tune: Raise on fast, cheap models to finish sooner.

Write a deduplicated curated snapshot every N batches.

tune: Raise to trade snapshot freshness for speed.

Per-chapter excerpt length persisted into the findings file.

tune: Raise for richer citations in findings.