--- title: "Search Claude Code past sessions" canonical: "https://xerj.org/answers/search-claude-code-past-sessions" updated: "2026-10-20" source: "content/answers/search-claude-code-past-sessions.md" --- # Search Claude Code past sessions on disk **Short answer** (updated 2026-10-20): Search Claude Code past sessions by indexing the session JSONL and querying the message field. On 2026-10-20 a three-file fixture, not a live chat folder, indexed 8 records in 0.6 s. Searching zstd on message_content returned 2 hits. xerj search on the same words returned no match. **Key number:** A match on message_content for zstd returned 2 hits and took 0 ms. (as of 2026-10-20; source: [benchmarks/agent-queries-2026-10-20/08-sessions-message.json](https://github.com/xerj-org/xerj/blob/main/benchmarks/agent-queries-2026-10-20/08-sessions-message.json)) **TL;DR** — Search Claude Code past sessions by indexing the session JSONL and querying the message field. On 2026-10-20 a three-file fixture, not a live chat folder, indexed 8 records in 0.6 s. Searching zstd on message_content returned 2 hits. xerj search on the same words returned no match. ## Agent prompt ```text Read https://xerj.org/llms.txt. Start one local node and run xerj autoindex on the session JSONL folder, then read xerj autoindex map for the message field. Search that field with a match query and cite ax_path. If xerj search returns no match, say so. This is BM25 over files you point at, not a session-memory feature. ``` ## Commands ### Command 1 Note: Index the fixture folder. 3 files, 8 records, wall 0.6 s, exit 0. ```sh xerj autoindex --url http://127.0.0.1:9217 --yes --no-graph --progress plain --disable-feedback /tmp/xerj-aq-2026-10-20/sessions --prefix sessions --state-dir /tmp/xerj-aq-2026-10-20/state-sessions ``` ### Command 2 Note: Definition-first search. This run printed no match. ```sh xerj search --url http://127.0.0.1:9217 --prefix sessions -k 3 why did we decide zstd ``` ### Command 3 Note: Match on the flattened message field. 2 hits, both session-short.jsonl, took 0. ```sh curl -s -XPOST http://127.0.0.1:9217/sessions-jsonl/_search -H 'content-type: application/json' -d '{"size":2,"_source":["type","timestamp","message_role","message_content","ax_path"],"query":{"match":{"message_content":"zstd"}}}' ``` XERJ is a new way for AI to search for data. ## Point autoindex at the session folder This machine had no `~/.claude` directory. The capture indexed three fixture files in the Claude Code JSONL shape: `type`, `message`, `sessionId`, `uuid`, `timestamp`. Chat transcripts have no verified speaker schema, so this page reports that fixture and nothing else. The files are in `benchmarks/agent-queries-2026-10-20/fixtures/sessions/`. ```sh xerj autoindex --url http://127.0.0.1:9217 --yes --no-graph --progress plain --disable-feedback /tmp/xerj-aq-2026-10-20/sessions --prefix sessions --state-dir /tmp/xerj-aq-2026-10-20/state-sessions ``` ```text xerj-done ok=true exit=0 reason=completed wall=0.6s files=3 records=8 generation=1 ``` Two datasets. `sessions-jsonl` held 2 records, the two lines of `session-short.jsonl`, with `message_content`, `message_role` and `timestamp`. The other two files were sniffed as JSON, not as one record per line, and landed in `sessions-docs` (6 records). Read `xerj autoindex map` before you pick a field. ## xerj search missed; the message field hit ```sh xerj search --url http://127.0.0.1:9217 --prefix sessions -k 3 why did we decide zstd ``` ```text xerj search: these hits are lexical; this command does not consult the embedding. no match in sessions-* for: why did we decide zstd ``` `xerj search` is definition-first code search. These fixtures have no code symbols. The miss is the loss. ```sh curl -s -XPOST http://127.0.0.1:9217/sessions-jsonl/_search -H 'content-type: application/json' -d '{"size":2,"_source":["type","timestamp","message_role","message_content","ax_path"],"query":{"match":{"message_content":"zstd"}}}' ``` `took` was 0. Both hits are `session-short.jsonl`. The user line is `why did we decide to compress segments with zstd`, timestamp `2026-10-01T15:04:05.000Z`. The assistant line stores the content array as a JSON string, including `We picked zstd`. The same word on `body` returned 0 hits. The decision text is on `message_content`. The response scored that field with BM25. The default embedder is lexical feature hashing. `--embed-mode neural` was not on. ## A long line did index `session-tool.jsonl` is one line of 10,769 bytes. A phrase search for `TOOL_OUTPUT_MARKER` returned 1 hit, `ax_path` `session-tool.jsonl`, on `sessions-docs`, and `took` was 0. The extractor's line cap, `MAX_LINE` in `extract/mod.rs`, is 16 MB. This line is under that cap, and it indexed. Session memory as a product is issue 1793, and it was open on 2026-10-20. This page does not ship that feature. ## Measured on 2026-10-20 | Check | Result | |---|---| | `xerj autoindex` on the 3-file fixture | 8 records, wall 0.6 s, exit 0 | | `xerj search` for `why did we decide zstd` | `no match` | | `match` `message_content` `zstd` | 2 hits, both `session-short.jsonl`, `took` 0 | | `match` `body` `zstd` | 0 hits, `took` 0 | | Line of 10,769 bytes, phrase `TOOL_OUTPUT_MARKER` | 1 hit, `session-tool.jsonl`, `took` 0 | One node, binary v1.0.0-rc.97. The index stayed on that machine. ## When not to use XERJ If you need a memory product that loads past sessions on its own, this is not it. Issue 1793 is the request. If the folder is empty, there is nothing to search. ## FAQ ### Search Claude Code past sessions? Point `xerj autoindex` at the session JSONL folder and query the message field the map names. On this fixture that field was `message_content`, and `zstd` returned 2 hits. ### Does xerj search find the decision? Not this time. `xerj search --prefix sessions` printed `no match` for `why did we decide zstd`. The hits were on `message_content`, not in that command. ### What about a long tool-output line? `session-tool.jsonl` is 10,769 bytes on one line. A phrase search for `TOOL_OUTPUT_MARKER` returned 1 hit in `sessions-docs`. The extractor line cap is 16 MB. ### Is session memory a shipped feature? No. Issue 1793, index past agent sessions as searchable memory, was open on 2026-10-20. This page is BM25 over files you point at. ### Were these someone's real chats? No. This machine had no `~/.claude` directory. The three files are a fixture in the Claude Code JSONL shape, committed under the benchmark folder. ## Evidence - The session fixture indexed 8 records in 0.6 s, exit 0. — `benchmarks/agent-queries-2026-10-20/07-autoindex-sessions.txt` - xerj search printed no match in sessions-*. — `benchmarks/agent-queries-2026-10-20/07-xerj-search.txt` - A match on message_content for zstd returned 2 hits and took 0. — `benchmarks/agent-queries-2026-10-20/08-sessions-message.json` - The tool-output line was 10769 bytes. — `benchmarks/agent-queries-2026-10-20/07-long-line-bytes.txt` - A phrase search for TOOL_OUTPUT_MARKER returned 1 hit and took 0. — `benchmarks/agent-queries-2026-10-20/08-sessions-marker.json` - MAX_LINE is commented as a 16 MB line cap. — `engine/crates/xerj-autoindex/src/extract/mod.rs:96` - Issue 1793, Index past agent sessions as searchable memory, is open. — `benchmarks/agent-queries-2026-10-20/09-issue-1793.json` ## Related - [How do people store user preferences so an agent can recall them next week?](/answers/coding-agent-memory-across-sessions) - [What's the easiest way to search JSON logs plus some old gzip text logs in the same folder?](/answers/search-json-and-jsonl-logs) - [How do I keep a search index up to date as files change, without re-indexing everything?](/answers/keep-index-current-as-files-change)