Transcript recall (a session searches its own earlier conversation)

What it is

A long session eventually fills its context window. When that happens, the AI's conversation is compacted: everything so far is replaced by a summary, and the detail is gone from what the AI can see — the exact error message, the contents of a file it read an hour ago, the output of a command.

That detail is not actually lost. The engine keeps its own full transcript on disk, including every tool call and everything each tool returned. Omniscio's own chat history does not keep tool output (it shows "result received"), so the transcript is the only place it lives.

Transcript recall lets the session search that transcript. Instead of re-reading files or re-running commands to get back something it already saw, the AI asks for the passages that contain the words it is looking for, and gets them back.

Where to find it

There is nothing to switch on and no screen to open — transcript recall is a tool for the AI, not a feature you operate. You meet it only indirectly: in what a session can still answer after a compaction, and in the hidden note the app hands each session when one ends.

How it behaves

How the AI finds out it can

The moment a compaction ends, Omniscio hands the session one short hidden note: its conversation was just compacted, its full earlier record is still searchable, and here is the exact command to run. You never see this note in the chat — it goes straight into the AI's context.

Every session gets exactly one of these after-compaction notes. A crew lead already receives its crew memory at that moment; the note rides along in the same message.

What it searches

• Everything the session has ever written to its transcript — what was typed to it, its replies, its reasoning, every tool call, and every tool's output.
• Every transcript file the session has used. After a compaction the engine often starts a new transcript file and leaves the old one on disk; Omniscio remembers each one and searches them all, newest first.
• Only that one session. A search never returns anything from another session's transcript.

How a search works

The AI calls GET /session/<its id>/search?q=<words> on the local control server with its own key.

• All words must match. Put a phrase in double quotes to match it as one piece.
• Up to 8 words, 200 characters. A blank or longer query is refused with a clear message.
• Newest matches first, 10 by default (&limit= up to 50).
• Each result is a window of text around the match — 2,000 characters by default (&maxChars= up to 20,000) — not the whole message, so a match deep inside a huge command output still comes back small. A cut end is marked …[cut].
• Each result says when it was written and what it was: something typed to the session, a reply, a tool call, a tool's output, or reasoning — and whether it came from a helper agent.

Safety

• Same access as reading the session's messages. A session's own key can search only its own record. The callers who may read another session's messages may search it too, and every such cross-session search is written to the access audit log.
• Secrets never come back. When Omniscio swaps one of your stored secrets into a command, the engine writes the real value into its transcript. Every passage is scrubbed of those values before it is matched or returned, so a search can neither find a secret nor show one.
• Read-only. Searching changes nothing.
• Never slows the app. Transcripts are read on a background worker thread, one search at a time. A very large transcript (100+ MB) takes seconds; if the search is busy, the AI is told to try again in a moment.

When there is nothing to search

If the session's transcript is not on this computer (for example it was cleaned up, or the session runs on an engine that does not write one), the search answers honestly with no results and says no transcript was found.

Not covered

• Sessions on engines other than the Claude command-line tool (DeepSeek and GLM sessions run through it, so they are covered).
• A helper agent's own internal steps, which the engine keeps in separate files; the helper's final answer to the session is in the main transcript and is searchable.

Related

Compaction summary explains what the AI keeps after a compaction, which is why recall exists at all. Session handoff is the other way to carry work forward — starting a fresh session that is already caught up — and Context warnings covers the 75% and 90% notes that come before a compaction.