nameforjt-afk/session-knowledge

What it solves

Claude Code's memory is scoped to individual project directories, meaning history from one project is inaccessible when working in another. This leads to redundant questions and the need to re-implement integrations or recall decisions made in previous sessions. session-knowledge solves this by indexing all local Claude Code session transcripts, making your entire history searchable and accessible across all projects.

How it works

The tool scans Claude Code's local .jsonl transcripts and creates a local index using SQLite. It splits the data into two paths:

1. Search Index: A redacted full-text search (FTS5) index of transcripts and instructions.
2. Credential Vault: A secure, restricted-access database (vault.db) that stores plaintext credentials scraped from sessions.

It provides 13 Model Context Protocol (MCP) tools that Claude can use to search sessions, reconstruct timelines, find previous tool calls (including failed ones), and retrieve credentials. It uses bigram pre-tokenization for CJK (Chinese, Japanese, Korean) language support and pattern-matching for secret redaction.

Who it’s for

Developers using Claude Code who want a global, cross-project memory for their AI assistant, allowing them to recall technical decisions, reuse code implementations, and manage credentials found in their session history.

Highlights

• Cross-Project Memory: Breaks the directory-level scoping of Claude Code's history.
• Privacy-First: Runs entirely locally using the Python standard library; no third-party dependencies, embeddings, or API calls.
• Secure Credential Handling: Separates redacted search text from a restricted-access vault for secrets.
• Implementation Discovery: Allows searching for code by "capability tags" (e.g., searching for "stripe" to find payment auth logic) rather than just symbol names.
• Error Recall: Ability to filter past tool calls to only those that failed to avoid repeating mistakes.