memex

name	memex
description	Use MemEX epistemic memory — record beliefs with trust/conviction/importance scores, recall them ranked, trace provenance, and keep contradictions and corrections as data instead of overwriting.
version	0.1.0
author	ai-2070
license	Apache-2.0
metadata	{"hermes":{"tags":["Memory","Epistemic Memory","Knowledge Graph","Provenance","MemEX"],"requires_tools":["recall","belief"]}}

MemEX — Epistemic Memory

MemEX is your long-term memory as a belief graph, not a notepad. Every belief carries who said it, why you believe it, what it conflicts with, and how confident you are. You interact through two tools:

• recall — retrieve beliefs ranked by trust × salience (not text similarity).
• belief — record and manage beliefs (remember, why, contradict, supersede, forget, stats).

The whole point is to preserve structure a flat note store throws away. Use it.

When to use

• Before answering anything about the user, their projects, prior decisions, or facts you may have stored → recall first. Do not answer from guesswork if memory might hold it.
• After learning a durable fact, preference, decision, or conclusion → belief remember, with scores set deliberately and parents linking the evidence it came from.
• When sources disagree → belief contradict (keep both; never silently pick one).
• When something is corrected, updated, restated, or overruled → belief supersede (replace without deleting).
• When asked "why do you think that?" → belief why to trace the support chain.

The three scores (the core idea)

Every belief has three independent 0..1 scores. They do not move together — set each on purpose. This is the cell flat memory can't represent (e.g. worth checking, not worth trusting).

Score	Question	Set it high when…
authority	How much should the system trust this, regardless of who said it?	primary/official source (filing, statute, the user stating a fact)
conviction	How sure was the author when they said it?	the source asserted it firmly / without hedging
importance	How much does this matter right now? (salience)	it drives a current decision or you'll need it soon

Quick reads of common combinations:

• High authority, high importance → a load-bearing fact. (audited number, user's hard requirement)
• Low authority, high importance → worth checking, not trusting — a rumor or unverified lead. Store it, flag it, verify later.
• High authority, low importance → true but currently irrelevant background.
• High authority, low conviction → a trusted source that hedged.

Belief fields (what each property does)

Field	Purpose
content	The belief itself, as a short statement.
scope	Namespacing key. Group related beliefs: user:<name>/profile, project:<name>, case:<id>. Recall can filter by exact scope or scope_prefix.
kind	What the belief is (see below). Default observation.
source_kind	How it got here (see below). Default agent_inferred.
parents	Ids of the beliefs this one was derived from — the provenance edge. Set this whenever a belief follows from others; it powers why.

kind — observation (something witnessed) · assertion (a claim) · assumption · hypothesis (unverified, may be wrong) · derivation (computed/reasoned from parents) · simulation · policy (a rule/standing instruction) · trait (a durable property of the user). Use policy/trait for things that should persist indefinitely; hypothesis for things you're not sure of yet.

source_kind — user_explicit (the user stated it) · observed (you saw it happen) · derived_deterministic (computed) · agent_inferred (you concluded it) · simulated · imported (from an external doc/source).

Tool: recall

Retrieve beliefs ranked by a weighted blend of the scores. Returns each belief's id, content, scores, and — when relevant — contradicted_by.

Param	Default	Notes
query	—	Keywords to match in content/scope (optional; omit for a pure ranked pull).
scope / scope_prefix	—	Restrict to a scope or scope tree.
kinds	—	Restrict to specific kinds.
min_authority	—	Drop beliefs below this trust floor.
weights	{authority:0.6, importance:0.4}	Raise importance for an attention queue; raise authority for high-trust recall.
contradictions	surface	surface returns both sides flagged with contradicted_by; filter keeps only the stronger side; none ignores conflicts.
decay	false	Down-weight stale beliefs by age.
limit	8	Max results.

// Start of a session: pull the user's most salient, fresh beliefs.
{ "scope_prefix": "user:laz/", "decay": true, "limit": 6 }

// "What's our deploy process?" — see disagreements, don't hide them.
{ "query": "deploy process", "contradictions": "surface" }

// Attention queue: low-trust but high-importance items worth verifying.
{ "weights": { "importance": 0.8, "authority": 0.2 }, "min_authority": 0.0, "limit": 5 }

Tool: belief

remember — store a belief

Set the scores deliberately and add parents for anything derived.

// Durable user facts the user stated → trait, high authority. Same trust, but
// importance differs by how much each will steer future answers — keep them
// distinct, don't make every preference 0.6.
{ "action": "remember", "content": "Vegetarian — no meat or fish",
  "kind": "trait", "scope": "user:laz/profile", "source_kind": "user_explicit",
  "authority": 0.95, "conviction": 0.95, "importance": 0.85 }   // a hard constraint
{ "action": "remember", "content": "Works as a structural engineer",
  "kind": "trait", "scope": "user:laz/profile", "source_kind": "user_explicit",
  "authority": 0.9, "importance": 0.6 }
{ "action": "remember", "content": "Prefers hiking holidays in the mountains to beach resorts",
  "kind": "trait", "scope": "user:laz/profile", "source_kind": "user_explicit",
  "authority": 0.9, "importance": 0.35 }                          // nice-to-know

// An unverified lead → low authority, high importance (worth checking).
{ "action": "remember", "content": "staging DB may be undersized for the new load",
  "kind": "hypothesis", "scope": "project:cyberdeck", "source_kind": "agent_inferred",
  "authority": 0.3, "importance": 0.85 }

// A conclusion you reasoned out → link the evidence via parents.
{ "action": "remember", "content": "the p99 regression came from the cache change",
  "kind": "derivation", "scope": "project:cyberdeck", "source_kind": "derived_deterministic",
  "parents": ["<trace-belief-id>", "<commit-belief-id>"], "authority": 0.8, "importance": 0.9 }

why — trace provenance

{ "action": "why", "id": "<belief-id>" }
// → { item, supported_by: [ ...the beliefs it was derived from... ], support_count }

contradict — record a conflict (keep both)

{ "action": "contradict", "id_a": "<belief-a>", "id_b": "<belief-b>",
  "rationale": "opposing readings of the same clause" }

Both beliefs stay live; future recall (in surface mode) flags each with contradicted_by.

supersede — a correction replaces the old belief without deleting it

Use for restatements, corrections, updated guidance, overrules. The loser's authority is lowered but it stays queryable ("what did we believe before?").

{ "action": "supersede", "winner_id": "<corrected-belief>", "loser_id": "<old-belief>",
  "reason": "user corrected the staging URL" }

forget — retract

{ "action": "forget", "id": "<belief-id>" }                 // single
{ "action": "forget", "id": "<root-id>", "cascade": true }  // also retract everything derived from it

stats — graph overview

{ "action": "stats" }   // counts by kind / source_kind / edge kind

Workflows

• New conversation → recall the user's scope to load context before responding.
• You learned something durable → remember it immediately (don't wait for session end). Pick the right kind/source_kind and scores; add parents if it followed from other beliefs.
• Two sources conflict → store both as separate beliefs, then contradict them. Surface the disagreement in your answer.
• Something changed/was corrected → remember the new version (with parents: [old-id]), then supersede old with new. Never edit-over the old belief.
• Asked to justify a claim → why on that belief and report the chain.

Worked examples

Illustrative scenarios — the point is the tool usage, not the facts. <...-id> is the id returned by an earlier remember.

Legal research (famous civil cases)

1 — Precedent + derived proposition (provenance). Store a binding precedent as a durable, high-authority belief; link the proposition you build on it via parents so why can justify the memo.

{ "action": "remember",
  "content": "Donoghue v. Stevenson [1932] AC 562 — manufacturer owes a duty of care to the ultimate consumer",
  "kind": "policy", "scope": "doctrine:negligence/duty-of-care",
  "source_kind": "imported", "authority": 0.95, "conviction": 0.95, "importance": 0.8 }

{ "action": "remember",
  "content": "client's contaminated-drink injury supports a duty-of-care claim against the bottler",
  "kind": "derivation", "scope": "case:acme-bottling",
  "source_kind": "agent_inferred", "parents": ["<donoghue-id>"],
  "authority": 0.6, "conviction": 0.7, "importance": 0.85 }
// Later: { "action": "why", "id": "<claim-id>" } → walks back to Donoghue.

2 — Supersession without deletion (point-in-time law). A later case overrules an earlier rule; the old rule must stay queryable for claims that arose while it controlled.

{ "action": "remember", "content": "Winterbottom v. Wright (1842) — no duty without privity of contract",
  "kind": "policy", "scope": "doctrine:product-liability/privity",
  "source_kind": "imported", "authority": 0.9, "importance": 0.4 }

{ "action": "remember", "content": "MacPherson v. Buick (1916) — manufacturer liable in negligence absent privity",
  "kind": "policy", "scope": "doctrine:product-liability/privity",
  "source_kind": "imported", "parents": ["<winterbottom-id>"], "authority": 0.95, "importance": 0.85 }

{ "action": "supersede", "winner_id": "<macpherson-id>", "loser_id": "<winterbottom-id>",
  "reason": "privity requirement abolished for negligence" }
// "current rule" → MacPherson; a 1900-claim query can still recall Winterbottom.

3 — Preserved disagreement (majority vs dissent). Don't flatten a split into one answer — store both and contradict so a memo surfaces the tension.

{ "action": "remember", "content": "Palsgraf majority (Cardozo): duty runs only to foreseeable plaintiffs (zone of danger)",
  "kind": "assertion", "scope": "doctrine:negligence/proximate-cause",
  "source_kind": "imported", "authority": 0.9, "conviction": 0.9, "importance": 0.7 }
{ "action": "remember", "content": "Palsgraf dissent (Andrews): proximate cause is a policy line, not foreseeability",
  "kind": "assertion", "scope": "doctrine:negligence/proximate-cause",
  "source_kind": "imported", "authority": 0.6, "conviction": 0.8, "importance": 0.6 }
{ "action": "contradict", "id_a": "<majority-id>", "id_b": "<dissent-id>",
  "rationale": "competing tests for proximate cause" }
// recall(scope="doctrine:negligence/proximate-cause", contradictions="surface") shows both, flagged.

Geopolitical research (US–China trade war, 2024–25)

1 — Differential trust on one event. The same impact claim from an institution, a state ministry, and an unverified lead must not rank equally. Set authority/conviction/importance to the source, not the topic.

{ "action": "remember", "content": "IMF: 2025 tariff escalation trims global growth ~0.5pp",
  "kind": "observation", "scope": "geo:us-china/trade-war/impact",
  "source_kind": "imported", "authority": 0.85, "importance": 0.8 }
{ "action": "remember", "content": "MOFCOM: US tariffs will 'inevitably fail', minimal impact on China",
  "kind": "assertion", "scope": "geo:us-china/trade-war/impact",
  "source_kind": "imported", "authority": 0.5, "conviction": 0.9, "importance": 0.7 }
{ "action": "remember", "content": "rumor: a major exporter is rerouting via Vietnam to dodge tariffs",
  "kind": "hypothesis", "scope": "geo:us-china/trade-war/impact",
  "source_kind": "agent_inferred", "authority": 0.2, "importance": 0.85 }
// Briefing: recall(weights={authority:0.8, importance:0.2}). The rumor stays low —
// but the maintenance sweep flags it on the attention queue as worth checking.

2 — A fact that changes over time (supersession). A tariff rate is restated after negotiations; keep the old one for "as-of" analysis.

{ "action": "remember", "content": "US tariff on Chinese imports raised to 145% (Apr 2025)",
  "kind": "observation", "scope": "geo:us-china/trade-war/tariffs",
  "source_kind": "imported", "authority": 0.9, "importance": 0.95 }
{ "action": "remember", "content": "Geneva talks: US tariff cut to 30% for a 90-day pause (May 2025)",
  "kind": "observation", "scope": "geo:us-china/trade-war/tariffs",
  "source_kind": "imported", "parents": ["<apr-rate-id>"], "authority": 0.9, "importance": 0.95 }
{ "action": "supersede", "winner_id": "<may-rate-id>", "loser_id": "<apr-rate-id>",
  "reason": "Geneva 90-day tariff reduction" }
// "current tariff" → 30%; "rate as of April" → still finds 145%.

3 — Competing analog frames (hypotheses + contradiction + provenance). Strategic reads are conditional; hold the rival frames simultaneously, each anchored to its evidence.

{ "action": "remember", "content": "analog: 2025 rhymes with Smoot-Hawley 1930 — escalating retaliation, contraction risk",
  "kind": "hypothesis", "scope": "geo:us-china/trade-war/frame",
  "source_kind": "agent_inferred", "parents": ["<apr-rate-id>", "<imf-impact-id>"],
  "authority": 0.5, "conviction": 0.6, "importance": 0.85 }
{ "action": "remember", "content": "analog: closer to US–Japan 1980s — managed friction, negotiated settlement, no systemic break",
  "kind": "hypothesis", "scope": "geo:us-china/trade-war/frame",
  "source_kind": "agent_inferred", "parents": ["<may-rate-id>"],
  "authority": 0.5, "conviction": 0.55, "importance": 0.85 }
{ "action": "contradict", "id_a": "<smoot-id>", "id_b": "<usjapan-id>",
  "rationale": "rupture vs managed-friction framing" }
// As the Geneva pause holds, the US–Japan frame gains weight — re-score it deliberately
// (remember the stronger read + supersede), never silently delete the rival frame.

Reacting to the maintenance block

Between turns, MemEX runs a cheap, read-only sweep and may inject a ## Memory maintenance (MemEX) block into your context. It is a to-do list — act on it:

• ⚠ stale (a belief whose parents were retracted) → its support is gone. Re-derive it from current evidence, or belief forget it.
• ⚠ unresolved contradiction → resolve it with belief supersede (winner replaces loser) once you know which side holds, or leave it and surface the disagreement to the user.
• 🔎 verify (low-trust, high-importance) → these are worth checking. Confirm or refute, then update: raise authority via a fresh remember+supersede, or forget if false.

The sweep never changes scores on its own — it only points; you decide.

Pitfalls

• Don't overwrite — supersede. Being corrected/restated is not the same as being wrong; the old belief is part of the record.
• Don't default every score to 0.7. Undifferentiated scores defeat the entire system. A user requirement and a hunch must not rank the same.
• Recall before asserting from memory. If the user asks about themselves or past work, recall first.
• Use scopes consistently so scope_prefix recall works (user:<name>/..., project:<name>, case:<id>).
• Link parents on derived beliefs — without them, why and cascade cleanup can't follow the chain.
• Low authority ≠ ignore. Low-authority/high-importance is worth checking; store it and verify, don't discard it.

Verification

After remember, confirm the belief is retrievable and scored as intended:

{ "query": "<a keyword from the content>", "limit": 3 }

Use belief stats to sanity-check the graph is growing as expected (counts by kind/source).