wiki-vault/Sources/Dev/ai-memory-architecture-research.md

created	path	project	status	tags	type	updated
2026-05-04	Sources/Dev	ai-memory-architecture-research	active	homelab ai memory architecture research	session-notes	2026-05-04

AI Memory Architecture Research

Session exploring the landscape of AI agent harnesses, model memory mechanisms, and persistent context layers. Started as a question about two products and turned into the architectural backbone for several follow-on projects.

Outcomes

• Distinguished harnesses from models as separate layers in the AI tooling stack. Harnesses are free or cheap; models are where the cost lives. Swapping harnesses doesn't require migrating data or model choice.
• Identified OB1 (by Nate B. Jones) as the right adoption target for a personal AI memory layer, validating the self-hosted MCP-server-plus-database pattern as the approach to project memory at scale.
• Mapped the conceptual gap between current production AI memory (transcript replay plus prefix caching) and the research frontier (soft prompts, persona embeddings, weight-level continual learning). Confirmed that frontier-API models will never expose substrate access; open-weight local models do.
• Reframed the original "build a harness" instinct as actually "build the memory layer." Most existing harnesses are good; what's missing is the persistent project memory underneath them.

Topics covered

Poolside

Frontier coding-focused AI lab, founded 2023 by Jason Warner (ex-GitHub CTO) and Eiso Kant. Up to ~$12B valuation after Nvidia's recent investment. Released the Laguna model family — including Laguna XS.2, a 33B-total/3B-activated MoE under Apache 2.0 open weights that runs on consumer hardware via Ollama. Their thesis: code is the best beachhead to AGI because it forces long-horizon reasoning.

Relevance: Laguna XS.2 is a candidate model for local agentic coding workloads on the RTX 3080 and slots cleanly into the Phase 2 OpenClaw plan.

Pi.dev

A minimal terminal coding agent harness by Mario Zechner (badlogic), MIT-licensed. Notable design choices: no MCP, no sub-agents, no plan mode, no built-in todos — extensions instead of features. Supports 15+ providers including Ollama. Tree-structured sessions with branching and gist sharing. Same category as Claude Code and Aider, different philosophy.

Relevance: Useful as a reference architecture for understanding what a harness is and isn't. Worth a look as a daily driver for local Ollama work, separate from sanctioned claude -p for scripted Claude usage.

OpenBrain (OB1)

Personal AI memory layer by Nate B. Jones. Self-hosted Postgres-plus-Supabase stack with MCP server, semantic search via pgvector, auto-extracted metadata, multiple capture sources (MCP, REST, Slack webhook, Obsidian import). Provider-agnostic — any MCP-compatible client can read and write through the same database.

This is the convergence point of the conversation. Independent implementation of the same architectural insight reached through reasoning: AI memory should be a separate, owned, queryable layer that any AI client can read and write through a standard interface, not a feature locked inside one vendor's product.

Adoption is the right call rather than building from scratch. The technical-project extension layer (project state, decision logs, session notes ingestion) is the gap to fill, and contributing it back is plausible.

Why this matters

The cliff-of-projects problem — what happens when there are 100 projects instead of 5 and Dispatch sessions can't carry coherent project memory — is solved at the architecture level by this category of solution. Not by a better harness, not by more discipline, but by a memory layer that lives alongside whatever tools come and go.

The research-frontier digression on soft prompts and weight-level memory established the boundary: production AI memory is RAG-shaped (text retrieved by vector similarity), not weight-shaped or activation-shaped. Frontier APIs will never expose the substrate. That clarifies what is and isn't worth chasing — pursue the practical memory layer (OB1), recognize the exotic stuff (persona embeddings, hypernetworks) as research-frontier and not currently buildable on closed-weight models.

Key learnings

• Harnesses and models are independent layers. Picking a harness and picking a model are separate decisions.
• "Memory" in current AI is the harness keeping a transcript and replaying it. The model itself is stateless. Prefix caching is a cost optimization, not a memory mechanism.
• Frontier API access stops at "send text, get text back." No weight access, no activation hooks, no soft-prompt injection. Open-weight local models are where substrate experimentation lives.
• Anthropic actively blocks third-party harnesses from using subscription OAuth (April 2026 enforcement). API key with workspace spend caps is the only sanctioned path for Claude access from non-Anthropic tools. claude -p headless mode is the loophole — invoking the official binary from a script is fine.
• The "build my own harness" instinct was actually about needing a shared memory layer underneath any harness. Different problem, different scope, much smaller project.

Follow-on projects

• Deploy OB1 on homelab — Supabase stack on Proxmox, MCP server reachable via Tailscale
• Wire Dispatch pre-compact hook to capture transcripts into OB1 via n8n
• Extend existing markdown summary n8n flow to also POST captures into OB1
• Build technical-project extension for OB1 (project state schema, decision log)
• Evaluate Laguna XS.2 on Ollama as local coding model
• Authelia in front of OB1 web surfaces (Studio UI), GoTrue stays for application identity

References

• Poolside Laguna: open weights via Ollama
• Pi.dev: github.com/badlogic/pi-mono
• OB1: github.com/NateBJones-Projects/OB1
• srnichols OpenBrain (alternative implementation): lighter, plain Postgres + pgvector, no Supabase dependency