Edit: github-backup/docs/10-memory-filter-and-digest-cadence.md

Note Content (Markdown)

# Memory Filter and Digest Cadence

## Decision

Hermes should not store every artifact it can see. The Easier brain should
store the business meaning of artifacts, with provenance and links back to the
source system.

Raw Fathom recordings, long transcripts, ad creatives, screenshots, exported
Google Drive files and similar bulky materials should remain in their source
systems by default. Hermes should ingest or reference them only through a
reviewed digest process.

This is not a small-storage preference. It is an agency operating principle:
agents have limited working memory, and bloated context makes them worse.

## Source References

The current decision pulls from:

- 8 Figure Agency Hermes guide:
  `https://leads.8figureagency.co/api/admin/upload/cmpih6o3h000012w4vpb7g90n`
- Dave Sifry metaswarm:
  `https://github.com/dsifry/metaswarm`
- Garry Tan gbrain:
  `https://github.com/garrytan/gbrain`
- GBrain agent install protocol:
  `https://raw.githubusercontent.com/garrytan/gbrain/master/INSTALL_FOR_AGENTS.md`
- Rhys Fisher's quiet-search critique:
  `https://thecognitiveshift.com/publications/let-them-run/your-ai-second-brain-has-a-quiet-search-problem/`
- Hermes official docs:
  `https://hermes-agent.nousresearch.com/docs/`

## Artifact Tiers

### Tier 0: External Raw Source

Examples:

- Fathom recordings and full transcripts.
- Google Drive creative files.
- Native Slack threads.
- Notion pages during migration.
- Airtable records while Airtable remains active.
- Easier Now product database records.

Storage:

- Remain in the source system.
- Hermes stores stable links, timestamps, owners and sensitivity labels.
- Pull only the slices needed for a digest or investigation.

Purpose:

- Evidence archive.
- Human review.
- Re-processing if a later tool is better.

### Tier 1: Ingest Receipt

Examples:

- `source_id`.
- source URL.
- capture date.
- participants or owners.
- sensitivity class.
- digest status.
- retry/error state.

Storage:

- Small Markdown page or structured row.
- Safe to keep in the server vault.

Purpose:

- Prove what was seen.
- Avoid duplicate ingestion.
- Support audits without loading the raw artifact.

### Tier 2: Working Digest

Examples:

- meeting summary;
- commitments;
- decisions;
- risks;
- unresolved questions;
- client relationship signals;
- SOP gaps;
- next actions.

Storage:

- Markdown in the vault.
- Cite Tier 0 or Tier 1 source.
- Default lifespan is short unless promoted.

Purpose:

- Daily and weekly operating review.
- Agent working context.
- Human review queue.

### Tier 3: Durable Knowledge

Examples:

- accepted SOP update;
- client/company profile fact;
- approved decision record;
- strategy principle;
- reusable prompt/skill;
- agent role update.

Storage:

- Compiled Markdown pages in canonical vault folders.
- Reviewed status and source references required.

Purpose:

- Long-term memory.
- Retrieval.
- Agent priming.

### Tier 4: Always-Loaded Context

Examples:

- `SOUL.md`;
- role-specific `soul.md`;
- hard gates;
- current weekly priorities;
- active-agent remit.

Storage:

- Tiny curated files.
- Strictly bounded and reviewed.

Purpose:

- Identity and operating constraints.
- Not a knowledge dump.

## Digest Rules

Every digest should answer:

- What changed?
- Who or what does it affect?
- What decision, commitment, risk or opportunity emerged?
- What source backs this?
- Is this durable knowledge, temporary working memory or noise?
- When should it expire or be reviewed?

Do not save:

- repeated small talk;
- full transcripts unless explicitly approved;
- whole creative files;
- channel history with no business event;
- model outputs with no review or source;
- duplicate summaries of the same source.

Preserve:

- exact commitments;
- pricing, scope, timeline and ownership changes;
- client sentiment and relationship risks;
- contradictions and dissent;
- strategy reasoning;
- why a decision was made, not just what was decided.

## Cadence

### On Event

Use when:

- a client call ends;
- a proposal or contract changes;
- spend, scope or timeline changes;
- an urgent relationship risk appears;
- a key internal decision is made.

Output:

- Tier 1 receipt.
- Tier 2 digest.
- optional approval request.

### Daily

The COO / AI Ops Manager reviews recent Tier 1 and Tier 2 items and creates:

- daily agency pulse;
- open loops;
- client risks;
- decisions needing Anthony;
- failed or stuck systems;
- next build priorities.

The daily digest should be short enough to read in 3-5 minutes.

### Weekly

The COO / AI Ops Manager creates:

- department scorecard;
- SOP gap list;
- agent performance notes;
- client relationship review;
- sales/marketing/fulfilment learning;
- durable knowledge promotions;
- archive/decay proposals.

The weekly loop decides what graduates from Tier 2 to Tier 3.

### Monthly

The system should run a memory hygiene review:

- stale pages;
- contradictory facts;
- unreviewed digests;
- orphan notes;
- high-volume sources creating low value;
- retrieval benchmark failures.

## Promotion and Decay

A digest becomes durable knowledge only if one of these is true:

- it changes how Easier should operate;
- it represents a client, prospect, team or product fact likely to matter later;
- it records a decision or commitment;
- it improves an SOP, scorecard, prompt, gate or agent profile.

Otherwise:

- keep it in working memory for a review period;
- link it to source;
- let it decay or archive.

## First Implementation

Before connecting live Fathom, Slack or Drive:

1. Create synthetic Tier 0 examples.
2. Produce Tier 1 receipts and Tier 2 digests manually.
3. Ask the COO / AI Ops Manager to run a daily review.
4. Promote only 3-5 durable notes.
5. Score retrieval against the promoted notes.

Only then connect real read-only sources.