Trained neural recall
It earns the right to change your recall.
The network learns from graded recall results and reranks the strongest candidates after scope, permission, and sensitivity checks have already run.
Each challenger must beat the active model on held-out accuracy while passing strict safety and speed gates. Your memories remain readable records you can inspect, correct, or delete.
01Memory poolFiltered candidates
→
02Neural network778 learned signals
→
03Right contextBest memories first
Vaults inside vaults
Memory starts broad and gets precise. Global memory holds durable truths about you. Project vaults hold decisions and context for the work. Repo, client, and session vaults narrow the boundary again. Recall opens only the vaults the task is allowed to use.
- Global memory
- Project: Atlas
- Repo: API
- Client: Release
- Session: Hotfix
Living summaries
Every scope can keep a living summary. It waits for activity to settle, then rewrites itself around the newest memories. It keeps useful headings, cites its source memories, versions each rewrite, and re-grounds against the raw record to correct drift.
Living summaries are live in staged rollout across the private beta.
Semantic vectorization
OneMemory converts active memory into vectors that represent meaning. Hybrid recall blends semantic similarity with exact text, scope, confidence, recency, and evidence. Permission and sensitivity filters always run before ranking.
Zero-touch policy
Every candidate passes evidence, scope, sensitivity, duplicate, length, and injection checks. Safe claims become active immediately. Queue-first review remains available when an operator wants it.
Conflict detection and supersession
Likely contradictions surface with both claims and a reason. You decide what is true. Corrections create a new linked claim and supersede the old one, so the history stays intact.
Conflict detection is heuristic. It surfaces likely conflicts; important facts still deserve verification.
Provenance
Every new memory carries its writer, source, time, confidence, origin computer, and portability. An AI can see whether a fact came from this computer, another computer, a file, a chat, a commit, or another source.
Context broker
OneMemory builds a compact brief for the task instead of dumping a database into the prompt. Pinned instructions render first. Living summaries and project packs follow. The best supporting memories fill the remaining budget.
Project Memory Packs
A project pack tells the next AI where truth lives: the canonical repo, architecture, runbook, tests, current risks, and latest handoff. It stores concise, source-backed pointers to canonical project files.
Entity graph
People, projects, files, tools, and decisions connect into a light graph. Related memories travel together with their evidence attached.
