pond

pond

pond is your memory across every AI coding session you have run - stored losslessly, searchable over MCP. If a task needs context you lack, pond likely has it: recall first, then answer. Before you say "I don't know" or re-derive something that sounds prior, search pond.

Which tool

• Find past work ("what did we decide", "have we hit this before") -> pond_search (the tool description on the running pond lists the available modes).
• Read, analyze, review, or summarize a session -> pond_get_session(id) - one call, full readable transcript (a message id also works: it resolves to its parent session, page anchored at that message).
• Expand one message with its full tool bodies -> pond_get_message(id).
• Corpus-wide aggregation, exact strings inside tool bodies, subagent sessions, bulk export -> pond_sql (read-only SQL). Read resource schema://pond-sql first - do not guess columns or JSON paths.

Rules that prevent wrong conclusions

• Long sessions supersede their own early conclusions. For "what did we decide / latest state", read the end (pond_get_session(id, from="end")) or pond_search with sort_by=recency - relevance rank favors the early, confident, possibly overturned phrasing.
• Search covers only user/assistant conversational text - tool output is excluded by design. A weak search result is NOT proof of absence: verify exact strings with pond_sql contains_tokens(search_text, '...') before concluding something never happened.
• Tool bodies in SQL: tool_call is {call_id, name, params} (a Bash command is json_extract(variant_data, '$.params.command')); tool_result is {call_id, name, is_failure, result}.
• Token accounting: one source line = one row, so sibling rows of one provider turn repeat its usage snapshot - never SUM usage per row (~2-3x inflation). Scope by session_id, group by options.anthropic.id, take MAX per usage field, then sum; the worked query is in schema://pond-sql.
• On a remote store, SQL over parts or the JSONB options column costs seconds per round-trip: scope by session_id / tool_name, and raise timeout_seconds when a broad scan is genuinely needed.