Personal search just drove into the agent bay.
Hister describes itself as a private search engine for pages you visit and files you keep. It indexes full contents, can ingest browser history and local directories, and exposes search through a web interface, terminal clients, and MCP. Its default posture is local or self-hosted with no telemetry or mandatory cloud service; optional semantic search can send document text to an embeddings endpoint chosen by the operator.[1]
That is useful because most context failure is not mystical. You saw the exact caveat three Tuesdays ago, buried in a release note, and now neither you nor the model can find it. Full-text recall turns “I vaguely remember” into a retrievable object.
A searchable pile is memory for finding. It is not memory for steering.
A transcript is cargo, not judgment.
Skillsync’s launch post makes a different move: translate agent sessions across incompatible on-disk formats. The authors say the system can carry messages, reasoning, and tool history between coding agents, while storing skills and memory as portable human-readable Markdown; session translation is local unless a user explicitly shares into a team workspace.[2]
VS Code 1.138 attacks continuity from inside the editor. Its Agent Host can keep the same session available across windows; the release also says Codex sessions can continue across apps, quick chats can attach to a folder or isolated worktree, and local sessions can run inside a Dev Container. Those are placement and continuity features—not proof that stale assumptions, permissions, or failed experiments should survive the move.[3]
A transcript tells the next driver what happened. It does not tell them which statement was later disproved, which tool output was environment-specific, or which permission should be revoked at the handoff.
Corrections need promotion, not accumulation.
Detail’s argument for “self-driving codebases” names global memory as a missing primitive: if one agent learns that an audit table is append-only, future agents touching that area should not need the same correction repeated. The post also argues that agent blind spots in development environments limit trust and that traces from bug-fixing work should expose which environment gaps deserve investment.[4]
The important word is not global. It is correction. A durable rule should be promoted because evidence changed the team’s operating model—not because a sentence happened to occur in a long chat. The promotion path needs provenance, scope, an owner, and an invalidation condition.
Build four racks, then put a gate between recall and action.
Corpus rack: pages, files, docs, release notes, and search indexes. Attach source, capture date, sensitivity, and retention. Session rack: prompts, tool calls, outputs, workspace identity, model or harness, and handoff summary. Preserve sequence without pretending every turn remains true.
Correction rack: reviewed lessons promoted into repository instructions, tests, runbooks, or policy—with an owner and expiry trigger. Authority rack: current credentials, tool grants, write targets, network reach, spend ceilings, and human brakes. This rack is checked live; it is not inherited from prose.
The salvage control below replaces a single “memory on” switch with four native intake gates. It executes no search and grants no permissions. Closing all four gates means the review template is ready to fill; the gauge measures declared sorting coverage, not truth, freshness, or safety.