Skip to main content
Each entry names the schema version served after the deploy, the date, and what a caller can see. Reads and writes that are not listed did not change.

0.3.26 (pending deploy)

  • POST /v1/forget takes space: {"tenant", "space"} deletes the blocks of one space of your tenant, the edges, verdicts and anchors that would otherwise point at an id with no block left, in one transaction. Fact declarations, backfill batches and blocks written without a space stay. A read of that space or of the whole tenant carries forgotten_spaces: [{space, at, rows}] in its certificate; a certificate with no forgotten space is unchanged. Without space the request forgets the whole tenant as before. See forget a tenant or a space.

0.3.25 (2026-11-20)

  • POST /v1/facts takes a key shape, */<rest>: one declaration such as {"fact_key": "*/language", "key_fields": ["subject"]} applies to every key shaped <subject>/language, so you no longer declare each subject’s key. An exact declaration of a key, in force at the instant asked, wins over the shape whenever it was declared. Any other * is refused with TCKG_BAD_FACT_PATTERN. resolve, the undeclared_fact write warning and measured strength all apply the shape. Export writes the shape out under each concrete key the bundle uses and carries no * key, so factblock validate accepts it. Keys declared exactly answer as before. See declare facts.

0.3.24 (2026-11-20)

  • Every read row carries fact_key and fact_value, the fact the block states a value for and the value as written, null for a block bound to no fact: GET /v1/memories, GET /v1/memories/{id}, search hits and why. You no longer need to copy them into payload to know what a row is the value of.
  • GET /v1/memories?fact_key_prefix= (and fact_key_prefix on the MCP tool read_memories) keeps the blocks whose fact_key starts with the string. With keys shaped <subject>/<attr>, fact_key_prefix=s_123/ reads one subject’s attributes at as_of and valid_at in one request. _ and % are plain characters. See list memories.
  • Rows of GET /v1/memories (including ?id= reads) and the item of GET /v1/memories/{id} carry superseded_by and superseded_since, decided exactly as on search hits at the same as_of, valid_at and facts. To learn whether a block you cited was later replaced, read it by id at a later as_of. Which rows a read returns and in what order is the same as before.

0.3.23 (2026-11-20)

  • Reading a few blocks by id (GET /v1/memories?id=, GET /v1/memories/{id}) and paging (limit, after) no longer read the whole space or tenant first. On a test space of 30,000 blocks an id read went from about 70 ms to under 1 ms, and a page of 100 from about 55 ms to about 14 ms. The rows, their order and the certificate are the same as before; nothing to change on your side.

0.3.22 (2026-11-19)

  • MCP track_record, timeline and graph run factblock 1.0.0a16 (was 1.0.0a14). track_record reads excess over any benchmark named in a verdict value ({"excess": ..., "benchmark": "069500"}), not only over SPY; existing keys are unchanged.
  • Nodes take asserted_at_precision, how precisely the time a statement was made is known: year, month, day, hour or minute, absent when exact. You can send it, or let the form of asserted_at say it: a bare date ("2026-11-19") is day precision and a time without seconds is minute precision. A value outside the list refuses that row with bad_precision. Every read that returns nodes (GET /v1/memories, GET /v1/memories/{id}, search hits, why) carries the field, null for an exact time, and export carries it on the nodes that have one. Nothing else changes: which rows a read returns and in what order is the same as before, and rows written earlier read as exact. See write memories.

0.3.21 (2026-11-19)

  • Reads take the case you are asking about: GET /v1/memories?facts= (a JSON object), POST /v1/search {"facts"} and POST /v1/resolve {"facts"}, and the same argument on the MCP tools read_memories, search and resolve. A block whose applies_if does not hold for the case is left out of the list, is neither a search candidate nor a hop, does not flag the block it supersedes, and is not a resolve candidate, so a SUPERSEDES edge from it is ignored too. Under most_specific a conditional finalist now answers for the case instead of needs_facts, which remains the answer when no facts are given. Every condition must hold; a condition on a key you did not give fails. Without facts every read answers exactly as before. The certificate carries not_applicable, the number of blocks set aside, whenever facts was given. See list memories, search and resolve.

0.3.20 (2026-11-17)

  • POST /v1/facts takes a fifth policy, most_specific: a candidate that a SUPERSEDES edge known by as_of and in force at valid_at points to is dropped, then the candidates with the most applies_if conditions are the finalists. Unconditional finalists answer when they all carry the same value and are unresolved_conflict otherwise. A conditional finalist is a new no_answer reason, needs_facts, because a read cannot pass facts yet. The four existing policies answer exactly as before. See declare facts and resolve.
  • POST /v1/search hits flag superseded_by only once the SUPERSEDES edge is in force at valid_at, not as soon as it is known, and carry a new field, superseded_since, the instant it took effect. A supersession announced ahead of its effective date no longer flags the block early. See search.

0.3.19 (2026-11-17)

  • Nodes take applies_if, the cases a block applies to as [{key, op, value}] with op one of eq, after, before. A malformed one refuses that row with bad_applies_if; the rest of the batch is written. The server stores and exports the conditions; it does not evaluate them yet. See write memories.
  • GET /v1/memories?fact_key= and POST /v1/search {"fact_key"} narrow a read to the blocks that state a value for one fact. Search narrows the candidates only, as category does. See list memories and search.
  • GET /v1/export carries applies_if on the node lines that have one (a bundle without it is byte-for-byte the same) and names its scope in two response headers, X-Tckg-Tenant and X-Tckg-Space. See export.

0.3.18 (2026-11-17)

  • GET /v1/edges and GET /v1/export take edges=touching with space: edges with exactly one end in the space are returned too, each with other_end: {side, id, space} (null when both ends are inside). Without the parameter the response is byte-for-byte the same as before. edges=touching without space is 400 TCKG_TOUCHING_NEEDS_SPACE; combined with with=strength it is 400 TCKG_TOUCHING_WITH_STRENGTH. The certificate counts touching edges in masked.edge. See edges and export.
  • A touching export is not a self-contained bundle: edges may point at nodes that are not in the file, so factblock validate reports dangling references. Use the default export for a bundle that stands alone.

0.3.17 (2026-11-05, API 2026-11-16)

  • GET /v1/verdicts: a target’s verdicts as of an instant, the latest chosen by the server, ties refused as unresolved_conflict. No more /v1/export round trip to read one verdict. See verdicts.
  • GET /v1/edges?with=strength on a space reads in about a second instead of 8 to 9 seconds (the edge read now filters by space before looking up nodes). Answers unchanged.
  • Reader queries are limited to 60 seconds; a read that exceeds it returns 503 TCKG_QUERY_TIMEOUT. Writes are not limited.

0.3.16 (2026-11-05)

  • POST /v1/forget {"tenant"}: an app deletes its own tenant’s rows in one transaction, with a record left behind; the certificate of a forgotten tenant carries forgotten: {at, rows}. See forget tenant.

0.3.15 (2026-11-05)

  • POST /v1/facts takes {"facts": [...]} to declare several facts in one request, each item reported as declared, unchanged or refused. See facts.
  • tckg.certificate with space reads the space’s node ids once; the API keeps one connection pool per role. Answers unchanged, large spaces read faster.

0.3.14 (2026-10-23)

  • New edge types SUPPLIES, COMPETES_WITH, OWNS, PART_OF (general family). Export declares non-core types in the manifest edge_types.