LoreConvo

LoreConvo v0.10.20

Your memory follows your identity, not your tool — with your consent.

LoreConvo is the only AI memory that carries your context across Claude Code, Cowork, Codex, Cursor, and Hermes Agent. One install, one memory, everywhere you code.

Install directly from Claude Code's plugin marketplace, or via PyPI: uvx --from loreconvo[server] loreconvo

Why LoreConvo?

Works wherever you work

LoreConvo works across Claude Code, Cowork, Cursor, Codex, and Hermes -- the same memory layer, no matter which client you reach for. When you switch mid-project, your context travels with you automatically.

Most tools wall off memory by machine or workspace. LoreConvo stores everything locally in a SQLite database you own, and surfaces it wherever you are. Capture happens two ways: explicitly via the save tools, or automatically at session end if you install the optional hooks. Either way, you can inspect, edit, or delete any memory at any time.

You control what gets saved

LoreConvo puts you in control: automatic capture only runs if you choose to install the session hooks, every save is inspectable, and you can delete any memory at any time.

Every memory shows you exactly where it came from — which surface captured it, when, what project context it belongs to, and which skill generated it. No mystery. Full provenance.

Your memory stays on your machine

LoreConvo stores everything in a SQLite database on your own machine. Your data stays local unless you explicitly enable optional AI summarization (off by default) by setting LORECONVO_ANTHROPIC_API_KEY, which sends a transcript excerpt or summary to the Anthropic API using that key. LoreConvo never reads the generic ANTHROPIC_API_KEY. No cloud accounts. No vendor with access to your session history.

Your sessions live in ~/.loreconvo/sessions.db -- a file you own, can back up, and can delete whenever you want.

Structured memory, not raw transcripts

LoreConvo captures two types of memory for each session:

• Episodic memory: what happened -- summaries, artifacts created, open questions left behind
• Semantic memory: what was decided -- stable conclusions about the project that persist across sessions

Together these give Claude a structured, searchable record of your project's history, not just a pile of chat transcripts.

Recall Benchmark

LoreConvo's FTS5 search is benchmarked against a 60-session synthetic corpus (6 topic areas, 36 labeled queries).

Compound token expansion (camelCase / snake_case query preprocessing) lifts Recall@5 by +35.7 pp on queries using technical identifiers like autoSave, pipeline_tracker, and get_context_for.

Full benchmark report | Reproduce

Quick Start

One command to install:

bash install.sh

This creates a virtual environment, installs dependencies, and verifies everything works. No system Python changes, no manual pip commands.

Using the Claude Agent SDK directly? git clone the public repo and point the SDK's local-directory plugin loader at it -- the repo root is a self-contained plugin directory (.claude-plugin/plugin.json + .mcp.json). No separate SDK-installable bundle exists or is needed.

Using LoreConvo

Claude Code (Terminal)

Start a session with the plugin loaded:

claude --plugin-dir /path/to/loreconvo

Or load it inside an existing session:

/plugin add /path/to/loreconvo

Replace /path/to/loreconvo with wherever you saved the source folder.

After making code changes, use /reload-plugins to refresh without restarting.

Once loaded, Claude has access to all 40 LoreConvo MCP tools automatically. Ask Claude to "save this session" or "recall what we discussed about X" and it will use the tools on its own.

Cowork (Desktop App)

1. Click the + button next to the prompt box
2. Select Plugins
3. Select Add plugin
4. Browse to the loreconvo source folder

Important: Shared Database Access

Cowork runs in a sandboxed VM and can't see your Mac's filesystem by default. To read sessions saved by Claude Code, ask Claude in Cowork:

"Mount my ~/.loreconvo folder"

Once mounted, Cowork reads and writes to the same database as Claude Code. Sessions saved in Code appear instantly in Cowork.

Claude Chat (Web)

Chat doesn't support plugins, so LoreConvo provides a one-command bridge. Run this in your terminal:

bash export-to-chat.sh

This exports your last session and copies it to your clipboard (macOS). Switch to Chat and paste (Cmd+V). Chat instantly has the context from your Code or Cowork session.

To search for a specific session:

bash export-to-chat.sh "tax prep"

How It Works Across Surfaces

The core value of LoreConvo is that context persists across Claude surfaces automatically. Here is the full chain:

Claude Code  (~/.claude/settings.json via `claude mcp add`)
  |-- SessionEnd hook --> auto_save.py --> ~/.loreconvo/sessions.db
  |-- SessionStart hook <-- auto_load.py <-+
                                           |
Cursor       (.cursor/mcp.json) <--MCP-----+
Codex        (~/.codex/config.toml) <--MCP-+
Hermes Agent (~/.hermes/config.yaml) <-MCP-+
Cowork       (MCP-native execution env) <--+
  All surfaces: save_session / get_recent_sessions / search_sessions

Claude Chat (web)
  |-- export-to-chat.sh --> clipboard --> paste into Chat

Claude Code is the primary surface. The hooks run automatically:

• When a session ends, auto_save.py captures the conversation and saves a structured summary (decisions, artifacts, open questions, tags) to the local SQLite database.
• When a new session starts, auto_load.py queries the database, scores recent sessions by signal quality, and injects the most relevant context into the session as system context. Sessions with open questions and decisions score highest; low-signal sessions are filtered out. It also indexes any MEMORY.md found in the project directory (see MEMORY.md Auto-Indexing below).

Cursor connects via .cursor/mcp.json in the project root -- the same MCP protocol as Claude Code. See INSTALL.md for setup details.

OpenAI Codex connects via ~/.codex/config.toml using a [mcp_servers.<name>] section. See INSTALL.md for setup details.

Hermes Agent connects via ~/.hermes/config.yaml under the mcp_servers: key. See INSTALL.md for setup details.

Claude Chat (web) does not support plugins. The export-to-chat.sh script bridges the gap: it exports your most recent session to your clipboard so you can paste it directly into Chat. This gives Chat the same context that Code would have loaded automatically.

The result: when you switch surfaces mid-project, you never have to re-explain what you were doing.

Your Data is Always Available

LoreConvo works through MCP tools when they are available and falls back to bundled scripts automatically when they are not. Your sessions are safe regardless of MCP status -- the same save, search, and recall operations work either way. You do not need to configure anything; the plugin skill handles the switch silently.

Project Workspaces

LoreConvo projects are persistent workspaces -- every session, decision, and artifact from your work on a project is searchable from any Claude surface.

# Create a project workspace
create_project("my-api", "REST API project", expected_skills=["openapi", "python"])

# Add persistent project instructions (optional)
create_project(
    "my-api",
    description="REST API project",
    instructions="Python 3.10+, SQLite only. No cloud dependencies. Deploy via Docker."
)

# See recent sessions, skill usage, and open questions for the project
get_project("my-api")

# Search scoped to the project
search_sessions("auth design", project="my-api")

Project Instructions (optional): When you create a project, you can store persistent instructions or constraints that Claude will see at session start. This is useful for enforcing project-wide standards without repeating them in every CLAUDE.md file. Instructions are displayed in the auto-load context before recent session summaries.

Used with LoreDocs, LoreConvo forms a portable project workspace for all of Claude -- session memory AND structured knowledge, entirely on your machine. Where cloud AI workspaces tie you to one ecosystem, the Lore pair works across every Claude surface you already use.

MEMORY.md Auto-Indexing

If your project has a MEMORY.md file, LoreConvo automatically indexes it at every session start. The contents become searchable alongside your regular sessions via search_sessions.

This means Claude can recall project conventions, team notes, or architectural decisions from MEMORY.md without you having to mention them. Search results from MEMORY.md are tagged memory_md and have source='file_memory' so you can tell them apart from regular session entries.

Which directory is scanned?

By default, LoreConvo scans the directory where Claude Code is running (the current working directory). To point it at a different directory, pass LORECONVO_PROJECT_PATH as an env flag in your claude mcp add --scope user command:

"--env=LORECONVO_PROJECT_PATH=/Users/YOUR_USERNAME/projects/my_project"

Replace YOUR_USERNAME and my_project with your actual values. Use the full absolute path -- do not use ~ or $HOME.

Filtering MEMORY.md entries in search results

To include MEMORY.md entries in a search, use search_sessions normally -- they appear automatically. To see only MEMORY.md entries, filter by tag:

"Search LoreConvo sessions tagged memory_md for 'database conventions'."

The index is updated each time a session starts (idempotent -- no duplicates accumulate).

Verify Installation

After installing, verify LoreConvo is working by asking Claude:

"Run get_recent_sessions and show me the results."

If you see a list of sessions (or an empty list if this is your first time), LoreConvo is connected. If you get an error about missing tools, re-run bash install.sh and reload the plugin.

For hooks verification (Claude Code only):

"Check if LoreConvo auto-loaded any context at the start of this session."

If the SessionStart hook is working, Claude will have received context from your recent sessions automatically.

Recommended CLAUDE.md Setup

For the best experience, add the following snippet to your ~/.claude/CLAUDE.md (global) or your project's CLAUDE.md. This tells Claude how to use LoreConvo consistently across sessions.

## LoreConvo (persistent session memory)

At session start:
1. Call `get_recent_sessions` to check for recent context relevant to the current work.
2. Use this context to avoid re-explaining things already discussed in prior sessions.

During the session:
- If important decisions are made or domain knowledge is shared, note it for the session summary.

At session end:
- Call `save_session` with a summary of what was accomplished, key decisions, open questions,
  and any artifacts created. Use appropriate tags (e.g., project name, surface).

For Cowork users: Cowork does not run hooks automatically. Add instructions to call get_recent_sessions at session start and save_session at session end in your project CLAUDE.md. See COWORK_RESTORE.md for details.

Plans: Free vs Pro

LoreConvo is local-first and free to use. Pro ($8/mo) removes the session limit and unlocks LLM-quality summaries, hybrid retrieval search, and cross-product linking. Everything runs on your machine on either plan -- Pro adds no cloud component.

Free tier search: keyword (FTS5) + recency ordering. Pro tier search: hybrid retrieval -- vector (BGE-small-en-v1.5), BM25 full-text, and recency reranking combined via RRF fusion. Finds sessions by meaning, not just keywords.

Check your current tier and usage with get_tier. Activate a Pro license with vault_set_tier.

Features

• Automatic session capture: Sessions save at session end and load at session start via Claude Code hooks -- no manual save_session call required
• Cross-client memory: Your context follows you across Claude Code, Cowork, Cursor, Codex, and Hermes -- not locked to one IDE or machine
• Structured sessions: Captures decisions, artifacts, open questions -- not just raw text; optional reasoning_notes field stores agent reasoning chains
• Project organization: Group sessions by project with expected skill sets
• Skill tracking: Record which skills were used for smart filtering
• Persona tagging: Hierarchical personas for agent-specific memory (e.g., ron-bot:sql)
• Full-text search: SQLite FTS5 for fast keyword search across all sessions
• MEMORY.md auto-indexing: Your project MEMORY.md is automatically indexed at session start and is searchable alongside regular sessions via search_sessions
• LLM async session summarization (Pro): Auto-saved sessions are upgraded to LLM-quality summaries in the background using Claude Haiku. Opt in by setting LORECONVO_ANTHROPIC_API_KEY. A daily cap (LORECONVO_SUMMARIZER_DAILY_CAP, default 100) prevents runaway API spend. Pro tier only.
• Embedding-based related session discovery (Pro): get_related_sessions automatically discovers sessions with similar content using BGE-small-en-v1.5 embeddings (cosine >= 0.75). Up to 10 bidirectional auto-links per save, same-project scoped. Free tier gets keyword co-occurrence links. Set LORECONVO_EMBEDDING_LINKS=0 to disable embedding links.
• Cross-product document linking (Pro): Automatically discovers and links the LoreDocs documents most relevant to any session, and vice versa. Uses two new tools: get_docs_for_session and session_link_doc. Requires both LoreConvo Pro and LoreDocs Pro.
• Configurable tool groups: Choose which groups of tools your assistant sees (core, memory, graph, agents, admin) with the configure_tools tool or LORECONVO_TOOL_GROUPS. All groups are on by default. See Tool Groups.
• Local-first: SQLite database, no cloud dependency, zero API costs

Tiers

Free - 50 sessions

• Full feature set: auto-load, full-text search, tagging, session linking, export and import
• Local SQLite storage -- your data, your machine, no cloud account required
• One-click install via the Anthropic Marketplace

Pro - $8/month

• Unlimited sessions
• Team memory: share sessions with teammates (local-first async export/import, no server required)
• Related session discovery and semantic search
• Anthropic managed-agents export
• LLM-quality session summarization in the background (async)

Upgrade to Pro -- $8/month

After checkout, your license key is emailed automatically to the address you used at checkout, usually within a few minutes. Questions: [email protected].

MCP Tools

LoreConvo provides 40 MCP tools that Claude calls automatically during sessions. The table below shows the most commonly used ones -- see MCP Tool Catalog for the complete reference.

Requirements

• Python 3.11+
• macOS or Linux
• mcp and click (auto-installed by install.sh)

Data and Privacy

LoreConvo is local-first. All data lives in ~/.loreconvo/sessions.db on your machine.

• Data collected: Session titles, summaries, tags, surface identifiers, project names, and skill names you provide when saving. No telemetry, usage analytics, or identifiers are collected automatically.
• Storage: SQLite database at ~/.loreconvo/sessions.db. No cloud storage. Override the path with the LORECONVO_DB environment variable.
• Third-party sharing: None by default. Data leaves your machine only if you enable optional AI summarization by setting LORECONVO_ANTHROPIC_API_KEY, which sends a bounded transcript excerpt or summary to the Anthropic API under that key. Every summarization path (background summaries (Pro), post-turn capture (Pro), and save_session(summarize=True)) uses only this key; the generic ANTHROPIC_API_KEY is never read. Leave it unset and everything stays local.
• Retention: Data is retained until you delete it via delete_session or remove the database file manually. No automatic expiry.
• Contact: [email protected]

Full privacy policy: https://labyrinthanalyticsconsulting.com/privacy

Troubleshooting

MCP tools not showing up in Claude Code? Make sure you ran bash install.sh first. The .venv must exist with dependencies installed.

"No module named 'mcp'" error? The .mcp.json points to .venv/bin/python3 inside the plugin folder. If you moved the folder, re-run bash install.sh.

Cowork can't see sessions saved in Code? Ask Claude to "mount my ~/.loreconvo folder" so Cowork can access the shared database.

Fallback Script (Direct DB Access)

If the MCP server is unreachable (e.g., in scheduled tasks or automation scripts), scripts/save_to_loreconvo.py provides the same core operations directly against the SQLite database.

# Save a session
python scripts/save_to_loreconvo.py \
    --title "Daily QA run" \
    --surface "qa" \
    --summary "Ran full test suite. All passing." \
    --tags '["qa", "automated"]'

# Read recent sessions
python scripts/save_to_loreconvo.py --read --limit 5

# Filter by surface
python scripts/save_to_loreconvo.py --read --surface code --limit 3

# Search sessions
python scripts/save_to_loreconvo.py --search "tax pipeline"

The script auto-discovers the database at ~/.loreconvo/sessions.db (or pass --db-path explicitly). It generates proper UUIDs and writes the same schema as the MCP save_session tool.

What's New

v0.10.20 (2026-10-26)

Added

• You can now choose which groups of LoreConvo tools your AI assistant sees. Fewer tools means less clutter in the assistant's tool list and less context used. Use the new configure_tools tool, the LORECONVO_TOOL_GROUPS environment variable, or a tool_groups setting in ~/.loreconvo/config.json. Changes apply after a restart. By default every group stays on, so nothing changes unless you opt in.

Fixed

• Related-session links created by earlier versions between sessions in different projects are now removed automatically on first start. v0.10.19 stopped new cross-project links; this cleans up the old ones. Links you created yourself are not touched. No action needed.
• Hermes users: LoreConvo tools that previously failed when called from Hermes now work, including export, onboarding, anti-pattern tagging and related-session lookup. Hermes now tells you when a tool is only available through the MCP server, and its built-in instructions use the current tool names.
• Pro: background session summaries now run when you have set LORECONVO_ANTHROPIC_API_KEY. They never started on databases set up by the normal install, so sessions kept their quick heuristic summary. If you set that key on Pro (for example only for summarize=True), LoreConvo will now also send new session text to the Anthropic API in the background to write summaries, up to 100 a day by default (LORECONVO_SUMMARIZER_DAILY_CAP). Unset the key to keep everything local.
• After your assistant closes normally, ~/.loreconvo/sessions.db now holds every session on its own, so copying that one file to another machine brings them with it. Previously, sessions saved through the MCP server could stay in a side file (sessions.db-wal), so a copy of sessions.db alone could arrive empty. If LoreConvo is still running or was force-killed, copy the sessions.db-wal file too.
• The Python fallback script now matches the MCP tools more closely: --search and --read hide external-tool and expired sessions by default (on --search, add --include-external or --include-expired to show them), --project now actually filters --search and --read (it was accepted and ignored; --inspect now rejects it with an error instead), --expires-at clear removes an expiry, and the script no longer erases the record that session-saving hooks are failing.
• python -m loreconvo.cli stats now shows whether hook saves are failing, like the get_stats tool.

Changed

• LoreConvo now requires Python 3.11 or newer. Python 3.10 reaches end of life this month, and the security fixes below need 3.11. If you are on 3.10, stay on v0.10.19 until you upgrade Python.
• Automatic summaries (save_session with summarize=True) and turn capture now use only LORECONVO_ANTHROPIC_API_KEY. Earlier versions also picked up a general ANTHROPIC_API_KEY set for other tools, which could send session text to the Anthropic API when you had not configured LoreConvo to do so. If you use either feature, set LORECONVO_ANTHROPIC_API_KEY. Without it, LoreConvo stays local-only.

Security

• Updated the embedding library and its image-handling dependency (Pillow) to versions that fix 37 published vulnerabilities. LoreConvo only embeds text, so the affected image code was not used, but the library was loaded.

See the full changelog for the complete release history.

License

Business Source License 1.1 (BSL 1.1) - Labyrinth Analytics Consulting

Free for personal/non-commercial use (up to 50 sessions). Commercial use requires a paid license. Converts to Apache 2.0 on 2030-03-31. See LICENSE for details.

mcp-name: io.github.labyrinth-analytics/loreconvo