pi-billion-memory is a local, read-mostly extension:
- It reads only the allow-listed compression files described in the README
(pi
.jsonl.acp.jsonsidecars and opencode-acpses_*.jsonstate files). - For an allow-listed pi sidecar, it may read the session file's first line
only (
cwd,id,timestamp,type,version) to resolve the project name. Message lines are never read. - It may open
opencode.dbread-only (orquery_only) only to map session IDs to working directories. Conversation content is never read. - It writes a local SQLite database, an optional log file, and the allow-list file under the user's home directory. It does not write to session files.
The extension makes no network requests, has no telemetry, and sends no data to
any service. The memory_search tool is a local database lookup.
Before a compression block is written to the SQLite store, its topic and
summary are scanned for common credential shapes: key/value assignments,
provider token prefixes, JWTs, PEM private keys, Authorization headers,
cookies, and Chinese credential labels. Credential values are replaced with
[REDACTED]. URLs (http/https/ftp/file/ssh/git/ws(s)/db schemes and bare
www. hosts) are replaced with [REDACTED_URL]; a hit count is logged without
the value.
This is a best-effort safety net, not a guarantee. Compression summaries should
not contain secrets in the first place; if you find a bypass, report it privately
as described below. The filter applies to newly ingested/updated blocks; existing
rows are not rewritten (delete ~/.pi/pi-billion-memory.db* and rescan to rebuild).
Please do not open a public issue for security or privacy problems. Use the repository's Security tab to open a private security advisory, or contact the maintainer through the GitHub profile linked from the repository.
Include:
- the affected version or commit;
- a minimal reproduction (neutral paths and fixture data only);
- the impact you believe is possible.
Do not include real session content, personal paths, tokens, or other users' data in the report.