Fred Memory and Agent Orchestration

Operating Model

Fred is the front door. GPT-5.5 reads the request, checks relevant memory, decides whether to delegate, reviews worker output, and owns anything user-visible or externally consequential.

Local models are not replacement brains. They are bounded task workers for draft reasoning, classification, code orientation, summaries, and first-pass checks on the Mac mini.

1prime orchestrator: GPT-5.5

4installed Ollama worker models

2queued optional worker models

Memory Orchestration

1
Runtime context first
Use the current conversation, workspace instructions, and active project files before rereading bootstrap material.
2
Recall only what is needed
Search memory before answering questions about prior work, decisions, dates, people, preferences, or todos.
3
Write raw events to daily memory
Use memory/YYYY-MM-DD.md for durable notes from the day.
4
Promote durable context carefully
Long-term facts go to MEMORY.md; project-specific state goes to the relevant project MEMORY_INDEX.md.
5
Refresh the semantic index
After material memory changes, run the OpenClaw memory index so recall stays useful.

Agent Orchestration

Stage	What Happens	Who Owns It	Output
Intake	Classify request, modality, project lane, privacy level, and whether it needs tools.	GPT-5.5	A scoped plan or direct action.
Context	Pull only the needed files, memory snippets, repo state, or project artifacts.	GPT-5.5	Task packet for Fred or a worker.
Delegation	Use a local model only for bounded work where draft output is enough.	GPT-5.5 routes; local model drafts.	Summary, classification, patch idea, or analysis.
Review	Check worker output against source files, user intent, and safety gates.	GPT-5.5	Accepted, revised, or discarded worker result.
Delivery	Send the answer, deploy the artifact, update memory, or ask for approval when needed.	GPT-5.5	User-visible final result.

GPT-5.5

Prime

Primary reasoning, orchestration, memory recall decisions, risk assessment, final review, and external-action control.

Never delegated: final product, architecture, security, release, messaging, money, privacy, legal, or public decisions.

qwen3:8b

Installed

Default broad local worker for project Q&A, synthesis, tool-prep, and draft reasoning.

Runtime note: use /no_think or a strict worker prompt for routine calls.

qwen2.5-coder:7b

Installed

Codebase exploration, small patch proposals, test triage, and implementation notes.

Guard: route by registry lane, not by model self-identification.

gemma3n:e4b

Installed

Cheap router, classifier, extractor, and lightweight summarizer for fast background checks.

Guard: ambiguous intent or user-visible output escalates back to GPT-5.5.

deepseek-r1:8b

Installed

Slower reasoning worker for debugging hypotheses, hard planning, and trade-off analysis.

Guard: it can analyze, but GPT-5.5 decides what action to take.

Queued Specialists

Pending

gemma3:4b is queued for first-pass vision and screenshot checks. granite3.3:8b is queued for business documents and structured extraction.

Status: Gemma pull was interrupted; Granite has not started.

Escalation Gates

A
User-visible output
GPT-5.5 reviews before anything is sent.
B
External action
Sending, posting, public deploy changes, broad imports, and syncs stay approval-gated.
C
Sensitive domains
Architecture, auth, releases, legal, finance, privacy, and relationship context stay under GPT-5.5 control.

Runtime Budgets

Rule	Default
Routine context	8K-16K tokens
Upper routine context for 8B models	32K only when needed
128K contexts	Controlled experiments only on this 16GB Mac
Parallel local models	Avoid multiple 8B workers unless memory pressure is low