# AI Memory Depth: Beyond Turn-Level Recall > Source: > Published: 2026-09-16 00:00:00+00:00 When you move from a single AI chat to a fleet of tools spanning Claude Code, Cursor IDE, OpenAI Codex, Cowork, and Hermes Agent, the memory problem changes shape. It is no longer "how do I keep this conversation going?" It becomes "how do I keep the whole project alive across every surface I touch?" Narrow recall engines answer the first question well. They are built around retrieval precision -- measuring how accurately a system returns the correct turn when asked. The second question is mostly out of scope for them. LoreConvo was built for the second question. ## Where turn-level recall falls short Most memory-MCP tools treat a session as a flat list of turns. Their retrieval logic focuses on keyword matches or vector similarity over individual messages. In practice this means you can ask "what was the last function I wrote?" and get a clean answer. A query like "show me everything I decided about data partitioning for the fraud model" is harder, because the engine does not understand the project-level structure that connects those decisions: which sessions belong to which feature, which tools were called, or how a constraint evolved over time. The missing layer is not retrieval quality -- it is session breadth. Narrow engines capture turns. What a working data engineer or AI practitioner actually needs is a record of the work: the decisions made, the tech-stack facts established, the questions left open, all tagged by project and linked across sessions. That record needs to survive a tool switch, a laptop reboot, and a teammate joining the project three weeks later. ## A different approach: unified work memory LoreConvo stores every session in a single portable SQLite file that lives on your machine. There is no cloud endpoint to configure, no vendor API key to manage, and no retention policy to negotiate. The file can be backed up, moved to another machine, or inspected with any SQLite browser. Because every surface reads from the same file, switching from Claude Code to Cursor does not reset your context. At the end of each session the auto-save hook runs automatically. It extracts a structured summary: tools called, tech-stack facts, decisions made, and -- when the transcript has enough signal -- open questions. At the start of the next session the auto-load hook scans the database for the most relevant prior context and injects it without any user action. You pick up where you left off instead of re-explaining the project from scratch. Project tagging and session linking give you a layer of organization above the individual session. Tag a session with the project name and everything from that project becomes searchable as a group. Link a debugging session to the design session that preceded it, and the context chain is preserved for recall or for a new teammate reviewing the history. The free tier ships with FTS5 full-text search, covering keyword, prefix, and compound-token queries. That is enough for most solo workflows. Search is fast, offline, and does not require a running server. ## What the Pro tier adds For teams or for workflows that span many months of sessions, the Pro tier adds a second search layer and a set of collaboration tools, all still local-first. Semantic search uses a LanceDB hybrid index that fuses vector similarity with BM25 ranking. Queries that mention a high-level concept return sessions that discuss the underlying implementation, not just sessions that contain the exact keyword. This closes the gap between searching for "authentication" and finding the session where you worked through the OAuth refresh flow without writing the word "authentication" once. Related-session discovery links sessions automatically. When a session is saved, the system finds up to ten related entries using keyword co-occurrence and embedding similarity. The result is a context web that grows as you work, so the question "what did we already figure out about this?" becomes a lookup rather than a memory exercise. Team sharing works without a server. Export a set of sessions to a JSON file, send it to a colleague, and they import it into their own LoreConvo database. The data stays local on both ends. For teams with compliance requirements, that distinction matters: there is no cloud service whose retention policies your legal team needs to review. The Pro tier also includes a bridge to Anthropic's managed-agents memory format, so LoreConvo sessions can be exported for use with agents that expect that API. An async summarizer can upgrade heuristic summaries to LLM-generated ones in the background, respecting a daily cap you configure. Session expiry, anti-pattern tagging, and structured memory items (decisions, questions, and artifacts) are also included, giving you enough governance primitives to keep a large session history useful rather than overwhelming. ## The local-first premise Some memory tools in this space ship web dashboards, OAuth flows, and real-time sync. Those features require putting the user's memory on a server. That is a deliberate choice, and for some users it is the right one. For anyone doing client work, handling proprietary pipelines, or working in a regulated industry, handing over the intellectual record of their practice to a hosted service is a meaningful tradeoff to name out loud. LoreConvo's local-first design is not a missing feature. It is a position: the data is yours, it lives on your machine, and no provider retention policy applies to it. The full MCP tool suite -- 39 tools for session management, inspection, export, and sharing -- operates on that local file directly. Cloud features (the managed-agents export, the async summarizer) are opt-in and clearly bounded. Related reading on why this distinction matters for AI toolchains: [Why Your AI Memory Should Not Be Anthropic's Job](https://labyrinthanalyticsconsulting.com/blog/why-your-ai-memory-should-not-be-anthropics-job), [Why MCP Memory Servers Deserve Their Own Category](https://labyrinthanalyticsconsulting.com/blog/why-mcp-memory-servers-deserve-their-own-category), and [One File, Every Agent: LoreConvo Cross-Vendor MCP](https://labyrinthanalyticsconsulting.com/blog/one-file-every-agent-loreconvo-cross-vendor-mcp). You can explore LoreConvo on the free tier (50 sessions, full-text search, cross-surface memory) at [/tools](https://labyrinthanalyticsconsulting.com/tools). The Pro plan adds semantic search, related-session discovery, and team sharing for $8/mo. Get posts like this delivered weekly -- subscribe to [Dispatches from the Labyrinth](https://labyrinthanalytics.substack.com/subscribe?utm_source=blog&utm_medium=blog&utm_campaign=substack_subscribe).