docs(avaliacoes): register TencentDB Agent Memory evaluation — Discarded

[pvilarim]

Summary

Type E evaluation of TencentDB-Agent-Memory as a cross-session agent memory layer for the SDD stack. Decision: Discarded for sdd-kit, openspec/infra.md and the normative pipeline.

Docs-only change. No behavior, no kit payload, no infra.

Changes

File	Change
doc/avaliacoes/2026-08-07-tencentdb-agent-memory.md	New evaluation (TEMPLATE.md shape)
doc/avaliacoes/README.md	Index row, top of table
openspec/project.md	Matching Non-goal, next to the Headroom one

Rationale

• Provenance (R2/R3). TDAM distills conversations into LLM-inferred tiers (atomic facts → scenario blocks → user persona) and recalls them into context with no provenance channel. Recalled inference is indistinguishable from sources 1–6, so R2 ranking and R3 gap-marking ([NEEDS VERIFICATION]) become unenforceable. Same failure mechanism that discarded Headroom, aggravated — there the agent acted as if it had seen everything having seen part; here it acts as if it had a source when it has a statistical inference.
• Overlap. Its Code-Graph asset duplicates GitNexus; LLM-Wiki duplicates Graphify + openspec/specs/. Identical criterion to the code-review-graph discard (2026-08-02).
• Policy. The Skill asset auto-generates skills, conflicting with sdd-skill-guidance (normatively offer-only) and the auto-generated-block anti-pattern (guide §2.5.1).
• sdd-session-handoff. The spec mandates a fresh chat per phase and refusal of /opsx:apply in an explore chat. Implicit cross-session memory reintroduces from below the coupling the spec forbids from above, without phase boundaries.
• Profile. Upstream gains (−61.38% tokens, +51.52% relative pass rate) come from long tool-heavy agentic loops. This hub is DOCS_SPECS — expected gain here is nil.

Only Chat Memory maps to an uncovered gap, and that gap is filled deliberately by sdd-skill-guidance (agent offers, human decides, result is versioned).

Re-evaluation conditions

Recorded in the doc: first-class Claude Code/Cursor integration on main via MCP (not a traffic-intercepting proxy); a recall mode that marks provenance; scope restricted to Chat Memory in an APP/HYBRID target repo as a per-project module (G3 precedent), never kit payload. Code-Graph, LLM-Wiki and Skill are out permanently.

Independent of adoption: the symbolic short-term memory pattern (task state as Mermaid, verbose logs offloaded to refs/*.md behind node-id pointers) is flagged as borrowable as an SDD task-pattern with no runtime dependency.

Notes

• Two claims are marked [NEEDS VERIFICATION] in the doc: Claude Code support (announced in the repo description, no implementation found on main) and the ~16.7k star count (API returned 403 through the egress proxy).
• openspec validate --all --strict was launched locally but did not finish within the timeout due to registry connectivity through the proxy; the sdd-gates workflow is the authoritative run. No openspec/changes/ or openspec/specs/ files are touched by this diff.

---

Generated by Claude Code