Fable Brain

Past Chats Tools

4 min read

Past Chats Tools

conversation_search and recent_chats are claude.ai’s inline retrieval tools over past conversations, verified in the export’s 24-tool inventory, though their detailed trigger rules are not yet captured in this vault’s claim packs.

What it is

  • Two of the 24 inline tools the claude.ai harness defines in JSONSchema format; the export places them in the retrieval cluster alongside web_search, web_fetch, and image_search (System Prompt Export 2026-07, L1711-3392).
  • They give Claude on-demand access to prior conversations, complementing the ambient memory system that derives userMemories in the background.
  • They are attested only for the claude.ai harness as captured in July 2026; nothing in this pack attests them for the API or Claude Code.

How it works

  • The tool inventory opens with “Here are the functions available in JSONSchema format” (System Prompt Export 2026-07, L1711-3392) and both past chats tools load inline with the system prompt.
  • They are not deferred: unlike the Google Workspace suites listed by name only behind tool_search (System Prompt Export 2026-07, L3238-3283), their schemas are immediately callable.
  • Functionally they implement a just-in-time retrieval strategy in Anthropic’s context engineering terms: keep lightweight identifiers in context and load data at runtime through tools instead of preloading everything.
  • Retrieved past-conversation content enters the current window, where the memory chapter’s framing rules still govern wording; “You mentioned…” style phrases are allowed only when the person directly asks about Claude’s memory (System Prompt Export 2026-07, L251-276).
  • The surrounding memory system is disabled in Incognito Conversations (System Prompt Export 2026-07, L195-203).
  • Sensitivity rules travel with recall: “Claude NEVER references memories with sensitive or upsetting content” (System Prompt Export 2026-07, L211) unless the user raises it first; the same bar reasonably extends to resurfacing retrieved transcripts.

Past chats retrieval is memory as a tool. The export separates ambient userMemories from explicit transcript search, and the memory chapter's framing and sensitivity rules bridge both surfaces.

The export chapter reportedly specifies when Claude should call conversation_search versus recent_chats, for example keyword recall versus time-bounded browsing. Those trigger rules are not in this claim pack; only the tools' existence, inline status, and cluster placement are verified.

Whether Incognito Conversations also disable conversation_search and recent_chats is not attested in this pack; the verified claim covers the memory system, not the retrieval tools.

Best practice

  • Prefer just-in-time retrieval for history: lightweight identifiers in context, data loaded at runtime through tools. EVIDENCE-BASED
  • Reach for past chats tools when the user references a specific earlier conversation, and leave ambient recall to the memory system. PRACTITIONER
  • Respect memory framing when presenting retrieved chats: no observation verbs and no data-retrieval phrasing such as “Based on” plus memory terms. EVIDENCE-BASED
  • Budget retrieval output like any tool response; accuracy and recall degrade as token count grows, so pull excerpts rather than full transcripts. EVIDENCE-BASED
  • Verify the harness before assuming availability; the capture only proves claude.ai. PRACTITIONER
  • Apply the memory chapter’s sensitivity bar to retrieved transcripts and do not resurface upsetting past content the user has not raised. PRACTITIONER

Pitfalls

  • Assuming the export’s trigger rules are known; this pack verifies existence, not the usage policy.
  • Dumping entire past conversations into context invites context rot on long sessions.
  • Presenting retrieved chat content as user profile data trips the forbidden-phrases list from the memory chapter.
  • Expecting these tools on the API or in Claude Code without verification; they are claude.ai inline tools in the capture.
  • Confusing past chats retrieval with userMemories; one is on-demand search over transcripts, the other is background-derived summary content.
  • Substituting transcript recall for freshness; past chats tools search prior conversations, not the web, so they cannot answer what changed since the knowledge cutoff.

Sources

Next actions

  • Ingest the export’s past chats trigger rules into a line-referenced claim pack entry.
  • Probe conversation_search versus recent_chats behavior in a live claude.ai session and record findings.
  • Check whether Incognito Conversations disable the retrieval tools as well as memory.

Graph View

Backlinks