Memory

Session lifetime, fact extraction, and nightly memory consolidation. For the broader picture (files, layers, templates) see Memory System.

URL: /settings?tab=memory

The panel is grouped into three sections in pipeline order: Sessions → Fact extraction (per session) → Memory consolidation (nightly). File-based upload retention lives on the Agent tab.

Sessions

Session timeout

How many minutes of inactivity end a session. When a session ends, Axiom may summarize it into the current daily note (see below). Default: 30.

Range: 1 – 1440 (up to 24h).

{ "sessionTimeoutMinutes": 30 }

When a session ends, the agent generates a short activity-log entry and appends it to memory/daily/<date>.md.

Session Summary Provider

Which provider writes the end-of-session summary into memory/daily/<date>.md.

• Use active provider (default) — whatever is selected under Agent → Provider.
• Explicit override — pick a cheap, fast model here to avoid summarizing with your most expensive reasoning model.

{ "sessionSummaryProviderId": "openai:gpt-5.4-mini" }

Fact extraction

After a session ends with at least N messages, run the transcript through a small LLM and write durable facts into the memories table. The agent retrieves these via the search_memories tool — useful once memory grows past what fits in a single prompt.

Enable fact extraction

Master toggle for the per-session extraction job. Default: true.

{ "factExtraction": { "enabled": true } }

Fact extraction provider

Dedicated provider (defaults to the active chat provider). Use something small and fast — this runs after every non-trivial session.

{ "factExtraction": { "providerId": "openai:gpt-5.4-mini" } }

Minimum session messages

Sessions shorter than this are skipped (no useful facts to extract). Default: 3. Range: 1 – 100.

{ "factExtraction": { "minSessionMessages": 3 } }

Memory consolidation

Scheduled job that condenses the last N daily notes into durable entries in MEMORY.md and trims the daily files. Keeps the prompt small without losing context.

Enable consolidation

Master toggle for the nightly job. Default: true.

{ "memoryConsolidation": { "enabled": true } }

Run at hour

Hour of the day (0 – 23, in your configured timezone) to run the consolidation. Default: 3 (i.e. 3 AM local time).

{ "memoryConsolidation": { "runAtHour": 3 } }

Lookback days

How many days of daily notes the job reads and condenses on each run. Default: 3. Range: 1 – 30.

{ "memoryConsolidation": { "lookbackDays": 3 } }

Consolidation provider

Dedicated provider for consolidation. Defaults to the active chat provider; set something cheaper if consolidation runs often.

{ "memoryConsolidation": { "providerId": "openai:gpt-5.4-mini" } }

Consolidation rules

A card at the bottom of this section links to /data/config/CONSOLIDATION.md — the prompt that tells the consolidator how to condense (what to keep, what to drop, how to name entries).

Click Open editor to jump to the Agent Instructions page. See that page for the default template and tuning tips.