2026 · Platform · Sole designer & engineer (schema, walker, provider, pane)
Vector recall and a knowledge graph in one Postgres — the model can find what you said, and you can see the walk.
hermes-memory ↗Agent memory is usually a vector dump or a hidden graph. Hubs collapse into a few Title-Case 'Concepts,' sessions leak into each other, and you cannot tell whether injection came from this chat or someone else's nouns. I needed durable local recall that grows with real conversation, stays inspectable, and does not pretend to be a hosted brain.
Two stores, one semantic topology. Apache AGE (PG17) is episodic only — this session, this turn, order in time. Postgres is the manifold: global `noun` rows, per-turn passports on `memory_chunk_nodes`, bivalent `semantic_edge` poles (`mentions` = co-occurrence, not theme). ANN entry is `conversations.embedding` (Ollama `nomic-embed-text`, 768-d). One bounded walker scores both prefetch and `/search`: `score = (0.4·sim + 0.4·align·provenance + 0.2·decay) × (magnitude / 8)`. Fountain docks nouns onto their turns and draws teal mention threads. Loopback only (`127.0.0.1:5450` / `:7890`). MIT. Alpha.
Postgres + AGE + pgvector as one MemoryProvider; loopback; honest install (you set the password, you seed three facts).
AGE `Turn → ABOUT → Concept` was the wrong semantic layer; hubs and session bleed.
Spec → V9 `noun` / passports / `mentions`; AGE keeps only Session/Turn/NEXT.
Same beam for prefetch and Fountain; nouns sit on their turn; cockpit/ghost levels taken off the ship surface.
Merge to `main`; Garden live at the pane URL; production V9 still a human migrate.
Episodic flower in AGE; semantic switches in SQL. No dual Concept/ABOUT + noun graph for the same labels.
`noun.label` is global; two sessions can mention Postgres; they do not share the walk's provenance.
Prefetch, `/search`, and the pane use one 7-tuple beam; no second ranking in the UI.
Fountain is optional grammar. Vector + graph injection works if you never open the pane.
Persistent memory & context injection for Hermes Agent — pgvector for semantic recall + Apache AGE knowledge graph for entity expansion
Proves: I can build inspectable agent memory — episodic flower in AGE, semantic manifold in SQL, one walker for recall.
Shipped as `hermes-memory` on GitHub (PR #67): V9 schema, mention-order extractor, drain that stops conversation `Concept`/`ABOUT`, passport isolation tests, Garden-only pane. A real chat seeds the first graph; verify uses synthetics on purpose. Live DB migrate remains an ops step so nouns appear in production memory.
hybrid-age MemoryProvider with V9 schema, pgvector ANN, AGE Session/Turn flower, ordered mentions, bounded walker, and optional Fountain 3D inspector.
Memory provider: schema, walker, provider, pane, install CLI
Architecture table, quick start, configuration reference, troubleshooting
Nouns, mentions, and Garden pane design spec
Layer map, walker scoring, drain pipeline
3D Garden inspector for Session/Turn/Noun walks