Mads Thines launches LoreKit to give coding agents local, inspectable memory

Mads Thines, a Copenhagen-based designer and product engineer, launched LoreKit, an open-source memory layer that gives AI coding agents a record of earlier mistakes, fixes, and project-specific lessons, storing memories as readable Markdown files locally. The initial workflow runs entirely on a developer's machine, requiring no account or hosted database, and is designed to be inspectable and editable with familiar tools. LoreKit emerged from Thines' two years of building autonomous workflows for coding agents, addressing the problem of agents rediscovering the same facts and fixes after sessions ended.

Mads Thines https://github.com/mthines?ref=runtimewire , a Copenhagen-based designer and product engineer, has launched LoreKit https://www.lorekit.io/?ref=runtimewire , an open-source memory layer that gives AI coding agents a record of earlier mistakes, fixes and project-specific lessons. The initial workflow runs entirely on a developer's machine, storing memories as readable Markdown files rather than requiring an account or hosted database. Thines detailed the local setup in an August 15 blog post https://www.lorekit.io/blog/give-your-agent-a-memory?ref=runtimewire . The project grew out of roughly two years Thines spent building an autonomous workflow for coding agents. He found that agents repeatedly rediscovered the same environmental facts and debugging fixes after sessions ended. His first answer was a collection of local files. LoreKit emerged when he wanted those lessons to move between machines, teammates and continuous-integration jobs without tying the memory system to one agent host. Reddit launch post https://www.reddit.com/r/AI Agents/comments/1v9aexa/sharing something ive been working on lorekit/?ref=runtimewire Thines' route into agent infrastructure is grounded in product work rather than AI research. His public profile describes him as a coder and designer working at observability startup Dash0, while his LinkedIn lists earlier training and work in graphic design before he moved into frontend and product engineering. LoreKit reflects that background: the central bet is on a memory system developers can inspect and edit using tools they already understand. LinkedIn https://dk.linkedin.com/in/madsthines?utm source=openai&ref=runtimewire A memory is a file, not a hidden model process Developers install LoreKit's CLI and MCP server https://github.com/mthines/lorekit?ref=runtimewire with: npx @lorekit/cli install The installer adds three agent skills, an MCP server entry and lifecycle hooks for supported environments. LoreKit's homepage names Claude Code, Cursor and Codex, while the repository says any MCP-compatible client can use the same memory tools. LoreKit https://www.lorekit.io/?ref=runtimewire In local mode, the developer adds a .lorekit.json file that selects local storage and points the MCP entry to the CLI's local server. Memories then live under ~/.lorekit/ or a repository's .lorekit/ directory. Each file contains Markdown and YAML frontmatter with fields such as scope, key, timestamps and the number of times a lesson has been encountered. LoreKit launch post https://www.lorekit.io/blog/give-your-agent-a-memory?ref=runtimewire In the example Thines published, an integration test fails because its Postgres container is not running. A LoreKit hook suggests preserving the fix, but the software does not record the session on its own; the model must call memory.write . Thines' resulting memory tells a later session to start the database before treating ECONNREFUSED 5432 as evidence of a code defect. The example is a founder-produced demonstration rather than an independent test of the system. LoreKit launch post https://www.lorekit.io/blog/give-your-agent-a-memory?ref=runtimewire That design gives developers a visible audit trail. A memory can be opened, searched with grep , committed to a repository, edited or deleted. Thines also describes the entries as advisory observations. Deliberate rules still belong in files such as CLAUDE.md , where a human can review and version them. LoreKit holds lower-confidence operational knowledge that may eventually become a rule, expire or remain a useful warning. LoreKit launch post https://www.lorekit.io/blog/give-your-agent-a-memory?ref=runtimewire Local memory is the distribution wedge LoreKit's local mode lowers the commitment required to test agent memory. Developers can run the tool without authentication, network access or a new managed data store. That choice also gives Thines a path into teams that may resist sending repository context or internal debugging lessons to an external service. The hosted product turns personal notes into a shared system. According to Thines, users can create an account, generate an API key and direct the CLI to LoreKit's remote Postgres-backed store. The software continues to keep existing local files on disk, and its list command presents entries from both locations. Sharing an older local lesson with teammates or CI requires the user to push it through a separate migration command. LoreKit launch post https://www.lorekit.io/blog/give-your-agent-a-memory?ref=runtimewire Thines says the storage choice sits below LoreKit's retrieval logic. Scope precedence, ranking, de-duplication and context budgets are supposed to work consistently whether the software reads from a local directory or the hosted database. That architecture is described in LoreKit's launch material and has not been independently tested. LoreKit launch post https://www.lorekit.io/blog/give-your-agent-a-memory?ref=runtimewire Organizations can bind repository scopes to a shared store and assign viewer, member, administrator and owner roles. The free hosted tier described in the launch post supports up to 5,000 memories and 120 requests per minute. Read-only tokens can expose team lessons to CI without allowing the pipeline to alter them. Those limits and permissions are company-supplied product specifications. LoreKit launch post https://www.lorekit.io/blog/give-your-agent-a-memory?ref=runtimewire LoreKit makes a narrower memory claim Agent memory has become its own software category. Mem0 https://docs.mem0.ai/platform/overview?ref=runtimewire sells managed memory infrastructure with automated extraction, confidence and decay scoring, and conflict handling. Mem0 introduction https://mem0.ai/blog/introducing-mem0?ref=runtimewire , long-term memory guide https://mem0.ai/blog/how-to-create-ai-agents-with-long-term-memory?ref=runtimewire , and memory decay announcement https://mem0.ai/blog/introducing-memory-decay-in-mem0?ref=runtimewire Letta https://docs.letta.com/guides/core-concepts/memory/memory-blocks?ref=runtimewire , an agent development platform formerly called MemGPT, puts structured, agent-editable memory blocks directly into an agent's context. LangMem includes tools for extracting, consolidating and searching semantic, episodic and procedural memories within LangGraph-based applications. LangMem documentation https://langchain-ai.github.io/langmem/?ref=runtimewire LoreKit currently avoids much of that machinery. Retrieval is lexical, using recurrence, recency, scopes and word overlap rather than embeddings or semantic search. A paraphrased query can therefore miss a relevant lesson. Thines says he load-tested the system from six to 60,000 memories, but that performance figure is self-reported and does not establish retrieval quality. No independent retrieval benchmark is cited in the available material. Reddit discussion https://www.reddit.com/r/LLMDevs/comments/1vq3qvl/ive built a free open source local and remote/?ref=runtimewire The narrower design is central to LoreKit's appeal. It treats agent memory as accumulated project notes with explicit provenance and human control. That makes it easier to understand why a lesson appeared in context and easier to remove one that has become wrong. Conflicting and stale memories remain the hard problem. LoreKit supports branch-specific scopes and time-to-live values, allowing temporary notes to become invisible automatically when they expire. Thines still leaves decisions about correcting, promoting or removing lessons with the developer, an important constraint for a system that asks agents to learn from their own behavior. LoreKit launch post https://www.lorekit.io/blog/give-your-agent-a-memory?ref=runtimewire LoreKit is early, and its hosted service introduces questions that local Markdown does not, including how teams will evaluate sensitive memories and prevent a mistaken lesson from circulating through multiple agents. Thines' local-first approach gives developers a practical way to test the underlying idea before making that operational commitment. The first useful memory can be a file they can read.