Most AI sessions begin with a blank slate. You open Claude Code, pick up where you left off in a data pipeline, and spend the first few minutes re-explaining the schema decisions you made two days ago, the library you chose over an alternative, the edge case that came up in a review. That context existed somewhere -- in a previous conversation, in a comment you wrote, in your head -- but the model has none of it. Starting from zero is not a limitation of the models; it is a gap in the tools that sit around them.
LoreConvo fills that gap. It ships several interlocking mechanisms for context restoration, each targeting a different layer of the problem. This post walks through how they work in practice.
The auto-load and auto-save hooks
The most fundamental mechanism is the hook pair. When you start a new session on any supported surface -- Claude Code, Cowork, the Codex desktop app, Cursor IDE, or Hermes Agent -- the auto-load hook fires before your first message. It queries the local SQLite database for recent sessions tagged to your current project, ranks them by recency and relevance, and injects a compact summary into the session context. The session begins with prior decisions already in scope, not as an afterthought.
At the end of the session, the auto-save hook fires without any user action. It extracts a heuristic summary covering what was discussed, the decisions made, the tech-stack facts surfaced, and open questions the model may signal during the conversation, captured opportunistically rather than guaranteed on every save. That record becomes input for the next auto-load. The loop is intentional: each session feeds the next, building a running record without any manual overhead.
Both hooks ship on the free tier. You do not need a Pro subscription to start accumulating context. The hooks are also cross-surface, which means a session started in Claude Code produces a record that the auto-load hook can inject into a subsequent Cursor IDE session on the same project. The context follows the project, not the tool.
Memory digest injection
After a few sessions, the raw session history grows longer than is useful to inject in full. The memory digest addresses this. get_memory_digest retrieves the current consolidated digest -- a distilled summary of decisions, patterns, and facts extracted from recent sessions -- and injects it at session start alongside the raw session summary.
The digest functions as a longer-range view. Where a raw session record captures what happened in a specific conversation, the digest captures what is durable: the direction you chose, the constraints you established, the terminology you settled on. A project accumulates institutional knowledge as it grows, and the digest makes that knowledge available at the start of every session without requiring you to scroll through weeks of history.
The auto-load hook uses get_memory_digest automatically when a digest exists. You can also call it explicitly from any MCP-compatible surface if you want to inspect or reload the current digest mid-session.
Recall: compaction on demand
The consolidate_memories tool (surfaced as "Recall" in the UI) performs a deeper analysis across recent session history. Where the auto-save hook captures individual sessions heuristically, Recall reads across a configurable window, identifies repeated themes and decisions, resolves contradictions, and produces an updated digest. It is the mechanism that keeps the digest sharp as sessions accumulate.
Recall is available as a single MCP tool call on any supported surface. You do not need to manage a background job or wait for a scheduled run. Run it when you want to reset the digest after a sprint, after a major refactor, or when you sense the injected context has grown stale. The new digest takes effect on the next auto-load.
Dream log: compaction history
Running Recall repeatedly over the same history would produce diminishing returns. get_dream_log exposes the compaction history -- a timestamped list of recent consolidation events, the sessions included in each run, and the mode used. This transparency lets you verify that the current digest reflects the work you care about, identify whether a particular decision was captured, and understand how the digest evolved over time.
The dream log is not a debugging tool. It is part of how LoreConvo makes memory auditable. If the injected context surprises you or you want to know why a particular fact appears in your digest, the dream log shows you exactly which sessions contributed to it.
Free tier and cross-surface coverage
Context restoration -- auto-load hooks, auto-save hooks, and memory digest injection -- ships on the free tier with no session-count limits on the hooks themselves. (The free tier limits stored sessions to fifty; once you reach that limit, older sessions roll off. Pro removes the limit and adds features like the LLM-async summarizer and related-session discovery.)
Cross-surface coverage means the same session record created by a Claude Code conversation is available to the auto-load hook in a Cursor IDE or Hermes Agent session on the same project. You do not need to configure different memory backends for different tools. One .mcp.json file in your project directory wires every supported surface to the same local SQLite store.
If you have been accepting cold-start friction as inevitable, it is not. The hooks are a one-time setup. After that, context restoration happens automatically, whether or not you are thinking about it.
Ready to try it? Explore LoreConvo on the Anthropic marketplace or get posts like this weekly: Dispatches from the Labyrinth.
PS: We are accepting early access signups for a lifetime deal -- join the waitlist to be notified when it opens.
Related reading:
- MCP memory servers deserve their own category -- why persistent session memory is distinct enough to warrant its own MCP category
- Why your AI memory should not be Anthropic's job -- the case for local-first, user-owned session storage
- AI Memory for Claude: An Honest 4-Way Comparison -- how LoreConvo compares to Anthropic's memory primitive and other options