v13.0.0 — MCP 2026-07-28, and honest benchmarks

[13.0.0] — 2026-08-27 — MCP 2026-07-28, and honest benchmarks

Every install created since the MCP Python SDK went 2.0 was broken. The
2.x line dropped the @Server.list_tools() / @Server.call_tool() decorators
this server was built on, and the dependency was floored at mcp[cli]>=1.0.0,
so a fresh pip / uvx / npx / brew / docker install resolved 2.x and
died at import with AttributeError: 'Server' object has no attribute 'list_tools'. Existing installs kept working only because their pinned 1.x
never moved. If you installed after mcp 2.0 shipped, upgrade.

Fixing it turned out to be the whole story: once tools registered again, the
server was already speaking the stateless 2026-07-28 protocol, because the SDK
serves both eras from one process. This release makes that explicit, adds the
2026 result shapes, and then goes back over the benchmark numbers — where a
feedback loop had been quietly inflating them.

Fixed — MCP SDK 2.x (breaking for anyone on a fresh install)

• src/server.py registers tools/list and tools/call through whichever
  API the installed SDK exposes: the 1.x decorators, or 2.x's
  add_request_handler(method, params_model, handler). MCP_SDK_ERA reports
  which path is live.
• mcp[cli]>=1.9,<3 in both requirements.txt and pyproject.toml. The
  upper bound is the actual lesson: an unbounded floor is how 2.0 reached
  users unannounced.
• A tool call arriving before _bootstrap_session() crashed the transport —
  store.raw_append sat outside the try. Now guarded.
• An unknown tool name returned {"error": "Unknown tool"} with
  isError: false. It now raises, so clients see a real error result.

Added — protocol revision 2026-07-28

• Stateless era served end-to-end. tools/list, server/discover and
  tools/call all answer without an initialize handshake, with protocol
  metadata carried per-request in params._meta. The legacy handshake era
  keeps working from the same process, so clients on older SDKs are
  unaffected. tests/test_mcp_protocol_e2e.py drives both eras against a
  real subprocess.
• structuredContent on every JSON-answering tool. 2026-07-28 lets
  structured content be any JSON value, so clients receive the parsed object
  instead of re-parsing a string. Error results carry text only. No
  outputSchema is declared, so this is purely additive.
• Behaviour annotations on all 74 tools — readOnlyHint,
  destructiveHint, idempotentHint, openWorldHint. Clients use these to
  decide what runs without a confirmation prompt. 38 tools are read-only;
  memory_delete, memory_forget, memory_update, kg_invalidate_fact and
  the two rebuild tools are marked destructive. A test fails if a new tool
  ships unclassified.

Added — Claude Code plugin

• .claude-plugin/plugin.json, .claude-plugin/marketplace.json, .mcp.json
  and hooks/hooks.json: the MCP server, the memory-protocol skill and the
  seven capture hooks install in one step.

  /plugin marketplace add vbcherepanov/total-agent-memory
  /plugin install total-agent-memory@vbcherepanov
  

• bin/tam_plugin_bootstrap.py resolves a runnable server in order —
  an existing .venv, PATH, the plugin's own checkout, uvx/npx, and
  finally a venv it creates — then execs it, so the client always sees one
  process. The first branch touches no network.

• hooks/pre-edit.sh and hooks/on-bash-error.sh moved from
  examples/hooks/ into hooks/ so the plugin references one directory.
  install.sh copies the whole directory as before.

Fixed — benchmarks measured their own history

Recall.search bumps recall_count on every row it returns, and the scorer
adds recall_boost = min(0.3, recall_count * 0.05). Spaced repetition is
wanted in normal use and fatal for measurement: successive runs against one
database scored 0.547 → 0.565 → 0.588 → 0.607 R@5 without a line of retrieval
code changing.

• Recall.search(..., record_usage=False) opts out. Both benchmark runners
  and memory_explain_search now pass it — a diagnostic must not change what
  it diagnoses. A clean run and a re-run are now byte-identical.
• benchmarks/locomo_bench.py had categories 2 and 3 swapped (it printed
  "multi-hop" over the temporal numbers and vice versa), and never computed
  overall MRR. Verified against locomo10.json and corrected.

Added — BEAM (ICLR 2026)

• benchmarks/beam_bench.py scores retrieval on
  BEAM at the 100K / 500K / 1M
  scales across its ten memory abilities, using each probe's source_chat_ids
  as gold evidence. No LLM in the loop, so it is deterministic and free.
• Measured: R@5 0.575 at 100K (5,732 messages, p50 17.7 ms) and 0.490
  at 500K (38,058 messages, p50 58.5 ms). Contradiction resolution, knowledge
  update and temporal reasoning hold above 0.78 at both scales;
  instruction_following and event_ordering are near-zero at both, which is
  a statement about the primitive rather than the tuning.
• Ingest throughput halved between the two scales (25.6 → 10.8 msg/s), so the
  write path scales with store size. Not yet profiled; recorded rather than
  explained away.

Added — LongMemEval measures the product now

• The runner reimplemented its own BM25 / RRF / MMR / CrossEncoder stack, so
  the published 96.2% described an algorithm, not this software. A new
  --modes store ingests each question's haystack into a real Store and
  queries Recall.search — the exact path an agent takes — and is now the
  default mode. The reference implementations stay available for ablations,
  with the docstring saying plainly that full and store are not the same
  claim.

Fixed — the server was carrying ~450 MB it never used

Reported by d.snezhinskiy, who noticed an MCP server sitting at ~1.5 GB and
sent a patch.

• chromadb and sentence_transformers were imported at module scope in
  src/server.py. Both are fallback paths — when fastembed is healthy
  neither is used — and sentence_transformers pulls in torch. Availability is
  now decided by importlib.util.find_spec and the real import deferred to
  first use. Measured on this machine:

  	before	after
  import server	558 MB	116 MB
  serving, steady state	1367 MB	909 MB

  torch is no longer loaded at all unless the fallback is actually reached.

• A failed fastembed init used to be one log line, after which the server
  quietly fell back to sentence-transformers and gained ~400 MB. It now names
  the model cache and the memory cost, because the usual cause is macOS purging
  the system tmp dir where fastembed caches models by default.

• New TAM_MODEL_CACHE opts into a durable cache location. It is not the
  default: moving the cache orphans models the user already downloaded, so
  every install would re-fetch ~500 MB once and offline test runs would fail.
  The contributor's patch pinned it unconditionally; this ships the diagnosis
  loudly and leaves the move to the people who need it.

The patch's other half pinned mcp[cli]<2. That is the same bug this release
fixes, and pinning would cap every user at the 1.x SDK forever, so it is not
taken — but it independently confirmed the failure on a client machine, with
the symptom MCP error -32000: Connection closed.

Fixed — other

• tree-sitter-language-pack was in no requirements file. The README sold
  "AST codebase ingest, 9 languages" as a differentiator while every user's
  ingest_codebase silently degraded to whole-file chunks. Now a dependency
  (2 MB wheel).
• The enrichment worker shared the Store's sqlite connection.
  check_same_thread=False permits cross-thread reads; it does not make
  writes safe, because sqlite3's implicit transaction lives on the Connection.
  Concurrent writes interleaved into cannot start a transaction within a transaction, which is what made long ingests die. The worker now opens its
  own connection — which is what WAL mode is for.
• Migration 028 failed on every fresh database, not just some. Root cause
  reported by @juicetin in the "separate
  migration observation" section of #12: Store._migrate() added the
  subagent-lineage columns before _apply_sql_migrations() ran
  028_agent_lineage.sql, so 028 always hit duplicate column name: agent_id.
  Because SQLite has no ADD COLUMN IF NOT EXISTS and executescript is
  all-or-nothing, it aborted before its CREATE INDEX statements and was never
  recorded as applied — so it retried on every startup, forever.
  • The schema change now has exactly one owner: the SQL migration.
    _migrate() no longer duplicates it, and a test fails if any column is
    ever added by both Store._migrate() and a SQL migration again.
  • The runner additionally replays a duplicate-column script statement by
    statement, so databases already wedged by the old behaviour recover
    instead of needing the operator to notice.
• ai_layer/verifier.py hardcoded ~/.claude-memory/nli_calibration.json
  and stopped finding calibrations after the ~/.tam migration. It now
  follows the resolved memory dir, with the legacy path as a fallback.
• The graph node cache was invalidated on every write. _ensure_node
  dropped the whole graph_nodes name cache each time it created a node, and
  extract_and_link dropped it again at the end — so the 60s TTL never
  applied and the next save re-read the table. Created nodes are now inserted
  into the cache incrementally; the TTL still picks up writes from other
  processes. A clean A/B over 4,000 saves measured 103.0 vs 99.9 saves/s —
  about 3%, within noise at that size, since the re-read cost grows with the
  node table. It is reported as measured rather than as the fix for the ingest
  slowdown it was chasing: that slowdown (25.6 → 10.8 msg/s between a 5.7k
  and a 38k record store) is not explained by this and remains open. Both
  arms of the A/B degrade on the same curve.
• vocabularies/ and filters/ were missing from the wheel and the Docker
  image. Same shape as the migrations bug PR #12 fixed, and just as quiet:
  src/ resolves them as parent.parent / <dir>, so in a git checkout
  everything worked, while every pip / uvx / npx / docker install fell
  back to an empty tag vocabulary and turned every memory_save(filter=…)
  into a no-op. An installed wheel now loads all 54 canonical topics and 11
  filter configs. tests/test_wheel_contents.py builds the wheel, checks each
  asset directory, and greps src/ so a new sibling directory cannot go
  missing the same way.

Fixed — test suite

The suite was red on a clean checkout: 21 failed, 9 errors.

• Eight modules stub the LLM seam but still had to pass config.has_llm(),
  which probes for a live Ollama in auto mode — so they failed on any
  machine without one. New llm_enabled fixture; the stub is what those tests
  measure, so no network call should ever be attempted.
• benchmarks/data/ is gitignored, so the LoCoMo few-shot and NLI calibration
  tests failed instead of skipping on a fresh clone.
• Embedding tests asserted raw vectors; the production OpenAI path
  L2-normalises so cosine equals dot product.

1881 passing, 0 failing.