AI Agent Memory Systems - August 2026 Landscape¶
TencentDB Agent Memory¶
Tencent Cloud open-sourced TencentCloud/TencentDB-Agent-Memory (repository created April 2026, v2.0.3 released October 2026, roughly 30,600 stars). It is a team-level memory hub that turns conversations, documents and code into four reusable assets: Chat Memory, Skills, LLM-Wiki and a Code-Graph.
Numbers from the repository's own benchmark table, measured with OpenClaw:
- 61.38 percent fewer tokens (221.31M to 85.64M)
- 51.52 percent relative pass-rate gain on WideSearch (33 percent to 50 percent)
- PersonaMem accuracy 48 percent to 76 percent, a 59 percent relative gain
The architecture is a semantic pyramid rather than a flat vector pile: L0 Conversation (raw dialogue) to L1 Atom (atomic facts) to L2 Scenario (scene blocks) to L3 Persona (user profile). Raw evidence is persisted in databases for full-text retrieval, and the distilled persona and scene layers are stored as human-readable Markdown, so every abstraction has a drill-down path back to the source text.
It ships a Hermes plugin as well as the OpenClaw one: the memory_tencentdb provider installs from npm, or as a Docker image that bundles hermes-agent and the gateway.
How It Compares to Existing Memory Systems¶
Sibyl-Memory (CorpusIQ stack)¶
- SQLite-backed FTS5 search with four tiers (warm entities, hot state, cold journal, reference docs)
- 95.6 percent LongMemEval score - ranked second globally
- Survives restarts, crashes, and model switches
- CorpusIQ integration: canonical facts, metric specs, source-of-truth registry
Honcho (Nous Research)¶
- Session-based memory derivation with peer-aware context
- Dream cycles for memory consolidation
- Conversation-level granularity with semantic search
GBrain¶
- Knowledge graph layer with embedded search
- Session handoff pages for cross-session continuity
- Works alongside Sibyl and Honcho as the third memory tier
Mem0 / MemGPT¶
- Earlier generation cloud-dependent memory
- API-based retrieval with managed embedding stores
- Higher latency, lower privacy
What Tencent DB Changes¶
Tencent entering the local memory space validates the approach CorpusIQ has been using since June 2026. The architecture patterns are converging:
- Local-first storage - no customer data leaves the machine
- Vector + keyword hybrid retrieval - FTS5 plus embeddings
- Tiered recall - working memory, recent context, long-term archive
- Cross-session persistence - memory survives restarts
The 51.52 percent pass-rate gain lines up with what we see in practice: structured memory removes the repeated context recovery that burns turns. The 61.38 percent token reduction comes from the same place - smaller working contexts because the agent recalls the distilled layer instead of re-reading the transcript.
Competitive Positioning¶
Tencent DB Agent Memory is open source and developer-focused. It does not ship with: - Business tool connectors - Canonical fact registry - Source-of-truth manifest - Metric specification DSL - Cross-source validation
CorpusIQ combines the memory layer with the business data layer. The memory system remembers what the AI learned. The integration layer gives it live data to learn from.