Deterministic Consolidation

A cortex you never tend slowly fills with clutter: the same fact saved three times, near-identical notes from a re-ingested feed, memories left floating with nothing linked to them. Graphnosis doesn’t just store your memories — it keeps them indelible: permanent, well-connected, and ever more retrievable. A set of background passes runs on a schedule, autonomously, with no prompts and no buttons to press.

The guiding principle is strengthen, never weaken. A memory you add grows more confident, more connected, and more integrated over time — through use, through corroboration, and through consolidation. Nothing fades from disuse. The only things that ever lower a memory’s standing are explicit correctness events — a contradiction is detected, the memory is superseded, or you correct it — and those are audited and reversible, never silent decay.

The Deterministic Consolidation tab — the third tab in the main window — is where you see all of this.

The Deterministic Consolidation tab

It is a holistic view: it spans every engram in your cortex, not just the one selected in the dropdown above. It shows:

• Memory health — a retrieval-quality report: can you find what you need, is it trustworthy, is it well-integrated.
• Self-healing — how many duplicate memories Graphnosis has merged on its own.
• Insights — patterns, gaps, and opportunities surfaced from your knowledge (requires a local LLM).
• Goals & Strategic Thinking — strategic plans Graphnosis is tracking, with deadline awareness.
• Recent activity — a live feed of what the passes are doing, plus the schedule they run on.

A Scan now button forces a full pass on demand; otherwise everything runs on its own schedule.

Memory health

Memory health rewards retrieval quality, not raw size. It blends:

• Connectivity — what fraction of your memories are reachable within their engram (no orphans).
• Integration — how interlinked the Cortex is beyond raw structure: cross-engram links and inferred connections.
• Confidence — the average confidence of your memories. Under strengthen-only this should trend up.
• Coherence — high when there are no unresolved contradictions.
• Reinforcement — how actively your memory is being used and strengthened.
• Weight spread — a guard against saturation: it warns if every connection has become equally strong, which would flatten retrieval ranking.

A quiet, sparse, perfectly-accurate engram is healthy. The headline number is 0–100; it reads 0 until the first real score is calculated.

Connection reinforcement

Every time your AI recalls a set of memories together, Graphnosis notices. Memories that are recalled together have the connection between them strengthened — and if no connection existed yet, a repeatedly co-recalled pair earns a new one. This is the heart of Deterministic Consolidation: your memory adapts to how you actually use it, so the connections you lean on rank higher in future recalls.

Reinforcement is strengthen-only and saturating: a connection climbs toward — but never past — full strength, and strong connections plateau on their own. A connection that isn’t used is simply left untouched; it is never weakened. New connections enter at a moderate baseline, so the cortex always keeps a meaningful spread of strengths.

Self-healing — merging duplicates

The same fact often lands in your cortex more than once: you re-ingest a file, a connector pulls an RSS item that was already there, an AI saves a note that paraphrases one you already had. Left alone, duplicates dilute recall.

Graphnosis scans for near-duplicate memories in the background. When it finds two that are provably the same, it merges them automatically — no prompt:

• Identical text — the two memories say the same thing word for word, once web markup and spacing are normalized away.
• Fully contained — one memory’s wording is entirely contained within a longer one, so the shorter one is redundant.

The bar is deliberately high: a merge happens only when dropping one side provably loses no information. A merge is a soft-delete — the redundant memory is set aside, not destroyed, recorded in the op-log, and recoverable (see Recovery).

What it won’t merge on its own — your review queue

Plenty of pairs look alike but aren’t provably the same: one has a number the other doesn’t, one adds a “not”, they overlap only partially. Merging those could quietly lose or flip meaning, so Graphnosis never does it automatically.

Instead they surface in the Check-in tab under “Needs your review”, side by side, and you make the call: merge them, or keep both. Graphnosis heals what’s certain; you decide what’s ambiguous.

Weaving connections

Isolated memories — ones with nothing linked to them — are a weak spot. Alongside duplicate detection, the same scan looks for memories that are clearly related but genuinely distinct and weaves an automatic “related” connection between them. This is deterministic and conservative: only strong matches are linked, and an already well-connected memory is left alone so the graph doesn’t turn into noise.

Consolidation — the deep pass

Once a day, a deeper consolidation pass integrates and tidies the cortex. It is deterministic, and — like everything else here — it only ever adds or tidies, never weakens:

• Transitive inference — if A leads to B and B leads to C, Graphnosis infers the A→C connection. Each memory becomes more connected.
• Community detection — it reads how your memories cluster, feeding the Memory health metrics.
• Redundancy cleanup — it removes dead connections (edges left dangling to an already-deleted memory) and exact-duplicate parallel edges. It never removes a connection between two live memories.

Cross-engram connections

Your engrams aren’t islands. Graphnosis links memories across engrams when they share meaningful named entities or are highly similar in meaning — so a query about a topic in one engram can surface what you know about it in another. These cross-engram connections are reinforced by use, just like connections within an engram, and are stored encrypted alongside your cortex.

When a source moves between engrams, Graphnosis removes the old cross-engram links (they referenced the origin’s now-deleted node IDs) and immediately runs a fresh cross-engram linking pass once the content is re-ingested in the destination — so related memories in other engrams are re-connected right away, without waiting for the next scheduled 6-hour pass.

Archiving an engram leaves all cross-engram connections intact. Archive is purely a visibility toggle — the graph file, nodes, and connection store are untouched. Unarchiving restores the engram with all connections exactly as they were. Only a permanent Delete purges cross-engram connections.

Stale-entry cleanup runs automatically as part of every cross-engram linking pass. Any connection anchored to a soft-deleted node (from a forgotten or moved source) is pruned silently before new connections are discovered. This keeps the connection store lean without requiring any manual action — including cleaning up connections from sources moved before this automatic rebuild was introduced.

Memory decay — and why it no longer touches your memories

Earlier versions slowly decayed the confidence of memories you hadn’t recalled in a long time. Under Deterministic Consolidation, that no longer happens to anything you’ve added. A memory you deliberately saved — a file, a URL, a clip, a saved conversation — never loses confidence from disuse. The decay machinery remains only for a future ambient capture feature (unconfirmed, auto-captured content), which is the only thing that should ever be allowed to fade.

When it runs

Graphnosis runs these passes on a schedule so they never compete with what you’re doing:

Pass	How often
Duplicate scan + connection weaving	every 20 minutes
Connection reinforcement	every 30 minutes
Connection forming (conceptual)	every 45 minutes
Goal check	every 4 hours
Cross-engram linking	every 6 hours
Insights	every 6 hours
Consolidation	every 24 hours

The first sweep waits about 60 seconds after you unlock. A duplicate scan also runs a short while after you ingest a file. The Scan now button forces a full pass any time, and the current schedule is always shown at the bottom of the tab.

Standalone vs. a local LLM

Everything deterministic above works with no AI model installed — this is the default, Standalone mode: Memory health, the duplicate scan and auto-merge, connection weaving, connection reinforcement, cross-engram linking, consolidation, and goal deadline tracking.

Adding a local LLM — via Ollama, the same on-device model that also upgrades the correction flow — unlocks the passes that need real language judgment:

• Insights — patterns, gaps, and opportunities across your engrams.
• Connection forming — deeper conceptual links than the similarity-based passes can make.
• Second-opinion review of past merges.

A local LLM is optional and always runs on your own machine — nothing about upkeep ever sends your memory anywhere.