Personal Context Management¶
Personal Context Management is the operating layer that turns Adam's links, notes, news, conversations, and agent work into reusable context that improves decisions, delegation, briefings, and Claude sessions.
Why this replaces PKM¶
PKM asked: "Where should this note live?"
PCM asks: "How does this source change what Adam, Hermes, or another agent should do next?"
The goal is not a tidy vault. The goal is operating leverage: faster onboarding, better questions, sharper decisions, stronger agent outputs, and less repeated explanation.
Three-layer model¶
1. Raw source library¶
Immutable source captures: URLs, extracted articles, PDFs, YouTube transcripts, Slack dumps, pasted thoughts, and metadata.
Purpose: - Preserve provenance. - Allow re-processing when models or goals improve. - Keep full source text out of compact operating context unless needed.
Default location: ~/.hermes/context-inbox/ for intake and ~/hermes-vault/raw/ or another explicit raw library once productionised.
2. Synthesised context / wiki layer¶
Curated, linked notes in ~/hermes-vault/ that answer:
- What does this mean?
- What does it connect to?
- What should Adam do differently?
- Which existing plans/context packs should change?
Existing examples: - [[Agentic-Analytics-Engineering]] - [[AI-Agents-in-Data-Engineering]] - [[LocalStack-First-30-Days-Context-Layer-Framework]] - [[Parallel-Agent-Orchestration-Playbook]]
3. Portable operating context packs¶
Compact, purpose-built context files for external agents and work sessions. These are not archives; they are deployment artefacts.
Examples: - Claude/LocalStack onboarding pack. - Career strategy pack. - Hermes operating pack. - Tech radar pack.
Default location: ~/context-portfolio/packs/.
Intake lifecycle¶
capture β extract β classify β connect β promote / action / archive
- Capture β Adam drops a link, thought, PDF, video, or instruction with minimal friction.
- Extract β Hermes fetches article text/transcript/metadata where possible and preserves the raw source.
- Classify β assign intent, domain, sensitivity, actionability, durability, and destination.
- Connect β identify existing wiki/context pages and related decisions/projects.
- Promote/action/archive β choose one or more fates:
- brief only
- save raw source
- promote to wiki
- update existing context page
- create task/action
- add to a context pack
- ignore/dedup
Provenance rules¶
Every durable item should carry:
- source_url or source_ref
- captured_at
- extracted_at when full text/transcript is fetched
- content_sha256 for raw text where possible
- confidence: high | medium | low
- freshness: evergreen | current | time-sensitive | stale
- privacy: private | work-sensitive | public-source | exportable
- export_policy: never-export | summarise-only | pack-safe | public
Work-sensitive material must not flow into public prompts, blog posts, consultancy material, or non-company tooling.
Review cadence¶
Daily intake processing¶
Process new context-inbox items. Stay silent when items are captured cleanly and require no decision. Surface only failures, urgent actions, or high-value synthesis.
Tech-news promotion¶
High-signal tech news should not evaporate after Slack delivery. Each promoted story gets a context fate: save raw, update wiki, update pack, create action, or discard.
Weekly context review¶
Review new sources and recently updated pages: - promote repeated themes into durable notes - prune or mark stale pages - update Claude/LocalStack packs - create actions for high-leverage follow-up - report only the delta Adam needs
Operating principle¶
Do not optimise for note volume. Optimise for future leverage.
A source is valuable only if it helps Adam ask better questions, make better decisions, produce better work, or give another agent better context.
Related¶
- [[Hermes Memory Architecture]]
- [[Hermes Vision]]
- [[LocalStack-First-30-Days-Context-Layer-Framework]]
- [[Agentic-Analytics-Engineering]]
- [[Parallel-Agent-Orchestration-Playbook]]