What CLIO remembers

Use past-session search to find an earlier decision or result and return to the conversation that recorded it. What CLIO keeps about you across conversations (preferences, projects, accounts) is long-term memory. For example, search for Fernhill to find why a report says that a station was first recorded empty at 8:15.

For the main search menu, highlighted passage navigation, and the Fleet-backed CLIO Search assistant, see Search your work.

Memory retrieves what was recorded. Before reusing a finding, open its source message and check the dataset, interval and qualifications. Evidence and lineage explains how to inspect the files and actions behind it.

Find an earlier conversation

1. Open Settings → Memory and scroll to Past sessions. Choose the session under Search in session.

2. Enter a distinctive word or phrase under Find earlier evidence or decisions, then choose Search. Search starts within the selected session.

3. Read the matching excerpt and its session label. Use Open on the result to return to the matching message in the conversation.

4. To include other sessions in the same workspace, select Include related sessions in this workspace and submit the search again.

Open returns to the matching source message. Relevance 100% means the query term matched; it does not verify the claim. Click to enlarge.

Search matches words in retained conversation text, recorded reasoning and errors. Try the name, unit or term used in the original conversation if a paraphrase finds nothing. An empty result means no retained text matched the query; it does not establish that the work never happened.

Agent access uses Tasks

In the Tasks candidate, agents discover permitted sessions with query_tasks and inspect their transcripts, recorded summaries and linked artifacts with observe_tasks. See Session coordination. The same task surface lets a parent inspect a failed child’s work before messaging it to continue.

The session_search, session_read_summary and session_read_context_frame tools let an agent search and read earlier sessions the same way. Turn them off with Search past sessions on the memory page; a conversation started with memory off does not get them either.

Where does the session summary come from?

The tool called Read session summary builds a small overview directly from the retained conversation. It does not ask a model to write a new summary. It returns:

• The session title, state, workspace and timestamps.
• The retained message count and message identifiers.
• Text excerpts from the last five messages, up to 360 characters per message. A message with no text, reasoning or error contributes no excerpt.
• The most recent rollback information, including excluded message identifiers.

This is a recent-message overview. It can omit an important decision from the beginning of a long session. Search for that decision and open the matching message rather than assuming the overview covers the entire analysis.

The returned source label is gact_visible_transcript_summary. The tool’s recorded arguments and result are available through its information button; the excerpt includes the original message identifier and time.

Inspect an actual returned overview

The recorded tool result comes from the fictional refinement session used in the comparison guide. Its 12 retained messages produce three text excerpts from the last five messages, because two of those messages contain no qualifying text. The final excerpt retains the limitation about individual journeys.

Compaction creates a different summary

When a conversation becomes too large for the model’s working context, CLIO can compact older context. This creates a model-written summary to carry earlier work forward while keeping the context within its budget.

Record	Where it comes from	What to use it for
Read session summary	Session metadata and excerpts from the last five retained messages; no new model call.	Orient yourself before opening the conversation.
Compaction summary	A model call over the older context selected for replacement, using CLIO’s built-in compaction instructions.	Understand what was carried into the model’s continuing context.
Context frame	A saved description of the sources and items assembled for a turn.	Check which items were included, their source identifiers and the reasons recorded.

Compaction uses the session’s selected model. When a session has no separate model selection, it uses the active host model. An unavailable selected model or an empty response produces a reported failure; CLIO does not silently switch to another model to summarize that session.

Open Settings → Memory → Session retention to inspect context usage and Compaction history. A retained summary has a link back to its conversation. For additions to a turn’s context, expand the receipt in the transcript or choose Open exact details; see context inspection.

The compaction summary is a model interpretation. Check consequential details against the original messages, files and tool results. Compaction is not independent validation of the scientific finding.

When the context fills

A compaction never deletes anything. The full record of the conversation stays stored and searchable; only what the agent reads from then on changes.

When it happens. CLIO compacts automatically when the context reaches a share of the model’s context window, 85% by default (autocompact.pct in Configuration). Change it for one conversation in Observability → Context → Compaction settings: Automatic compaction turns it on or off, the slider sets the threshold from 50% to 95%, and Compact now compacts at once.

What the agent reads afterwards. Its instructions, then the summary, then what is kept word for word, then the steps that follow. By default CLIO keeps word for word:

• the messages you typed, newest first, as many as fit in 80,000 characters (compaction.keep.user_chars); a message too long for what is left is summarized, and older ones that still fit are kept;
• the request the agent is working on now;
• its newest step, so the output it was using stays exact;
• the session plan, its todo list and the conversation’s facts, such as its workspace folder, which are offered again.

The compaction.keep.* settings change these defaults.

The summary. The model that runs the conversation writes it, under fixed headings: the objective in your words, your requirements, progress, findings with their sources, decisions, files, open threads and the next step. The instructions for it are part of CLIO. The earlier compaction.prompt_file setting is no longer read: if it is still set, compaction fails and reports compaction_prompt_invalid, so remove it from your configuration.

When a compaction fails, or the context is still over the threshold just after one, the conversation shows a notice. The agent does not see these notices.

Recall what a summary replaced

The summary’s last line names its compaction. With recall_context, the agent gets back the steps a summary replaced, as they were recorded: your messages, messages and reports delivered to it, each tool call and each result, oldest first. A summary that replaced an earlier summary leads back to that one’s steps. Where CLIO’s storage can search, the agent can also search earlier steps for words and get the matching ones, best match first, at most 50.

Recall returns only what the agent’s own context would have shown: a result still waiting for its authorization reads as such, and input not yet given to the agent is left out. A recall result has a fixed size, 16,000 characters (tools.recall_result_chars). A longer result is saved to a file, and the agent gets its beginning and the file, which it can search and read on. Recall does not return the model’s reasoning.

Ask CLIO to recall earlier work

Make the source and scope explicit:

In this workspace, find the earlier conversation about Fernhill becoming empty.
Show the matching message and the interval it describes before reusing the finding.

The three memory tools come from the memory plugin. An agent can use them only when it carries that plugin. The standard agent does not, so first add memory to the conversation with the + menu.

Tool	What it actually returns
session_search	Ranked matching excerpts with session, message and part identifiers. Ranking measures matching query terms, not factual confidence.
session_read_summary	The recent-message overview described above.
session_read_context_frame	A saved frame’s metadata and up to 50 item descriptions, including source identifiers, inclusion and reasons. It is not the full text of every source.

Default tool calls stay within the current session. Workspace recall expands to related sessions in the same workspace. Global recall refers to sessions in CLIO’s global workspace; it does not mean every workspace on the computer. Read requests targeting another ordinary workspace are rejected.

Check what recall can leave out

Keyword retrieval can miss synonyms; a recent-message overview can miss older decisions; model-written compaction can omit details. Open the original message and supporting files before relying on a recalled claim.

Implementation references

• Summary projection and memory routes: constructs the overview from retained messages and records policy decisions.
• Transcript search: word matching, ranking and bounded excerpts.
• Agent memory tools: the callable tools, scope and supplied user-intent checks.
• Memory route tests and summary contract tests.
• Compaction model selection and its tests.

The service checks the scope and intent supplied by the calling tool. These fields are not a separate human approval mechanism.