Axis Bridge Solutions
← All capabilities
Implemented Maturity: advanced

Agent Memory & Knowledge Systems

Persistent, queryable memory architectures — from thread context to a self-hosted bitemporal knowledge store.

Executive summary

An agent without memory is a party trick. Axis Bridge designs memory as owned infrastructure: PostgreSQL as the system of record, embeddings and extraction as replaceable machinery around it, originals never mutated, and facts that keep their history instead of overwriting it.

What this capability means

Memory is where most agent projects quietly fail: context windows overflow, facts go stale, and nobody can say why the agent “remembers” something wrong. Axis Bridge treats memory as a data-engineering discipline. The store is owned (your PostgreSQL, your hardware, your backups), the pipeline around it is replaceable, and every derived artifact — extraction, embedding, fact — traces back to a preserved original through a provenance log.

The maturity ladder above is not theoretical: all three levels are running today, from the daily-use messaging assistant to a LAN-deployed knowledge system with enforced privacy tiers.

Maturity ladder

  1. 01

    Basic — Context injection

    ● demonstrated

    Static instructions plus recent conversation packed into the prompt window.

    User → Prompt context → LLM → Response
  2. 02

    Intermediate — Structured memory store

    ● demonstrated

    Durable stores the agent reads and writes — interaction logs, media descriptions, thread history across both sides of a conversation.

    Message → Memory store → Retriever → LLM → Reply + log
  3. 03

    Advanced — Governed knowledge system

    ● demonstrated

    A dedicated knowledge base with full-text documents, chunk embeddings, bitemporal append-only facts, provenance, audit events, and a review queue — served through one access-controlled doorway.

    Sources → Ingestion → Documents · chunks · facts → Guarded doorway → Any AI client

Evidence

  • Deployed six-table knowledge store — documents with full extracted text, ~512-token chunks with 1024-dim embeddings, subject/predicate/object facts, provenance, governance events, review queue. me-inc-os-knowledge / db/migrations (PostgreSQL 18 + pgvector, live on Coolify)
  • Bitemporal, append-only facts — outdated facts are closed with an end-timestamp, never edited, so "what was true in June" stays answerable forever. me-inc-os-knowledge / PLAN.md §6
  • Three durability commitments — Postgres as system of record, originals never mutated, privacy enforced at the doorway — written down and built to. me-inc-os-knowledge / README §1
  • Production messaging memory — every interaction logged, image and sticker content described by a vision model and embedded into thread history so later drafts have full context. me-inc-os-api / CHANGELOG.md, Session 43
  • Version-controlled knowledge packs synced into the LLM front-end on a 30-minute schedule, portable by design. llm-knowledge-base (Gitea) + oikb sync