What's New in MDDB 2.18.8

29 October 2026

MDDB 2.18.8 is for agents that work on code. An agent that edits a project spends most of its context finding things. It greps, opens a file, opens another, and reads hundreds of lines to change three. This release makes finding cheaper and easier to measure. It also stops two agents that share a document from losing each other's edits.

The full list is in the changelog. This article covers what changes for you.

A repository in one command #

mddb-cli autoindex .

This indexes every text file under the directory into one collection. It respects .gitignore and records each file's path, kind and language, and the server extracts symbols from code. A second run sends only what changed. On the MDDB repository, 1,455 files and 14 MB, the first run took 41 seconds and the second took half a second.

Passages instead of documents #

Any search can now answer with format: "passages". Each result is then a place: a file path, a line range, the function or type that holds those lines, and the lines themselves, with an estimate of their size in tokens. An agent asking "where is the revision computed" gets the dozen lines that do it, not the 400-line file around them.

For code collections the search advisor recommends passages. A collection can make them its default.

Symbols, not just text #

GET /v1/symbol and the MCP tool code_symbol answer everything about one name. They return where it is defined, with the span from start line to end line, and every place that uses it. They also return its callers and callees, the tests that touch it, and other definitions in the same directory. Before an agent changes a function, one call tells it what else will notice.

Code to learn from #

Before implementing something, it helps to see how a project that already did it went about it:

mddb-cli reference add github.com/etcd-io/bbolt --ref v1.4.0
mddb-cli reference search "how are freed pages released"

The CLI makes a shallow clone into your cache directory and indexes it as a reference collection. It records the repository, the commit and the licence. MCP agents search every reference collection at once with reference_search, reference_symbol, reference_examples and reference_dependencies. The server never clones anything. That stays on your machine. Reference coding has the details.

Narrower searches #

Filter operators. filter accepts prefix, contains, ranges (gt, gte, lt, lte), not and exists, alongside exact matches. Dates and numbers compare as dates and numbers. "Go files under services/, changed this week" is one search.
Grouping. groupBy returns at most k results per file, author or collection. A search no longer hands back seven chunks of the same long document.

Is it searchable yet? #

A write returns once its full text and metadata are indexed. Its embedding is computed in the background. Until now, an agent that wrote a note and searched for it by meaning a second later found nothing, and could not tell why. Searches now take waitForIndex and waitForKeys to wait for named documents. GET /v1/indexing-status and the MCP tool get_indexing_status report the queue. The SSE event collection.indexed fires when a collection's queue empties.

Fewer tools, chosen deliberately #

MDDB has 90 MCP tools, and every schema sits in the model's context on every turn. Toolsets let a server or a single API key expose only what an agent needs. MDDB_MCP_TOOLSETS=search,code cuts tools/list from about 19,600 tokens to about 8,900.

Edits that cannot overwrite each other #

Every document now carries a revision. Pass it back as expectedRevision, or expected_revision over MCP and gRPC, and the write fails with 409 when someone else wrote in between. The second agent then reads again instead of silently overwriting the first. Building this showed that REST, gRPC and MCP had three separate update implementations that had drifted apart. MCP updates, for example, skipped the metadata index. They now share one.

langchain-mddb uses this for MddbBackend, a filesystem backend for LangChain Deep Agents. Files live in MDDB with revision history, they can be searched by text and by meaning, and concurrent edits are safe.

Numbers instead of guesses #

Two features let you measure what search does instead of trusting it.

Evaluation sets. Store the questions people actually ask, each with the documents that should answer it. Then compare configurations:

mddb-cli eval docs --config fts/bm25 --config hybrid --config vector

The output gives recall@k, nDCG@k, MRR, latency and tokens for each configuration. With a set stored, the search advisor recommends from the measurement instead of from the collection's shape. --fail-below recall@10=0.8 turns it into a CI gate. On MDDB's own documentation, with 20 questions, hybrid search ranked best: nDCG@10 0.783, against 0.739 for BM25.

Retrieval gain. The server now counts the tokens its searches return. It also counts the tokens that passages and chunks save compared with the whole documents they came from. mddb-cli gain reports both over a window of days, and so do GET /v1/gain, the panel's System page and Prometheus. The counters keep counts only, never queries.

We also measured an agent finding code in this repository, on 15 tasks. Grep followed by reading the right file reached 13 answers with about 140,000 tokens. One vector search returning passages reached the same 13 with about 85,000 tokens, in half the calls. When the agent reads the files with the most grep matches first, as agents usually do, grep needed 2.8 million tokens. Full-text passages reached only 6 of the 15. These questions are phrased the way people ask, and the code uses other words. That is the kind of thing an evaluation set on your own questions will show you. The benchmark has the method and every number.

Aliases and bulk changes #

An alias is a stable name for a collection. Build docs-v2 with a new embedding model, then run mddb-cli reindex-swap docs --into docs-v2. It copies the documents, waits until they are indexed, and moves docs to the new collection in one step. The old collection stays, so a rollback is one command.

delete_where, update_where and set_ttl_where act on every document a metadata filter matches. Each document goes through the normal write path, and a dry run shows first what would change.

gRPC catches up #

The gRPC search calls were separate copies of the search code, and they had fallen behind REST and MCP. They now call the same code, so gRPC gets filters, grouping, passages, read-your-writes, symbol lookup, evaluation and gain as well.

Fixes worth knowing about #

Two writes in the same second no longer lose a revision (#358).
Raising test coverage to 97% for every file this release touched found several crashes. A cancelled gRPC batch, a short memory session ID, a missing MCP-keys bucket, and startup with automations disabled could all crash the server.

Upgrading #

Nothing to migrate. All new fields are optional, and a write without expectedRevision behaves as before.

One status code changes. A request for a document or revision that doesn't exist now answers 404 instead of 400, with the same JSON body. A client that checks for 400 to detect a missing document has to check for 404.

Download 2.18.8 from GitHub releases, or run mddb-cli self-update.