Files
hermes-agent/plugins/memory/holographic/retrieval.py
Alexander Whitestone aa6eabb816
Some checks failed
Forge CI / smoke-and-build (pull_request) Failing after 23s
feat(memory): enable temporal decay with access-recency boost
The holographic retriever had temporal decay implemented but disabled
(half_life=0). All facts scored equally regardless of age — a 2-year-old
fact about a deprecated tool scored the same as yesterday's deployment
config.

This commit:
1. Changes default temporal_decay_half_life from 0 to 60 days
   - 60 days: facts lose half their relevance every 2 months
   - Configurable via config.yaml: plugins.hermes-memory-store.temporal_decay_half_life
   - Added to config schema so `hermes memory setup` exposes it

2. Adds access-recency boost to search scoring
   - Facts accessed within 1 half-life get up to 1.5x boost on their decay factor
   - Boost tapers linearly from 1.5 (just accessed) to 1.0 (1 half-life ago)
   - Capped at 1.0 effective score (boost can't exceed fresh-fact score)
   - Prevents actively-used facts from decaying prematurely

3. Scoring pipeline: score = relevance * trust * decay * min(1.0, access_boost)
   - Fresh facts: decay=1.0, boost≈1.5 → score unchanged
   - 60-day-old, recently accessed: decay=0.5, boost≈1.25 → score=0.625
   - 60-day-old, not accessed: decay=0.5, boost=1.0 → score=0.5
   - 120-day-old, not accessed: decay=0.25, boost=1.0 → score=0.25

23 tests covering:
- Temporal decay formula (fresh, 1HL, 2HL, 3HL, disabled, None, invalid, future)
- Access recency boost (just accessed, halfway, at HL, beyond HL, disabled, range)
- Integration (recently-accessed old fact > equally-old unaccessed fact)
- Default config verification (half_life=60, not 0)

Fixes #241
2026-04-13 15:38:12 -04:00

640 lines
23 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
"""Hybrid keyword/BM25 retrieval for the memory store.
Ported from KIK memory_agent.py — combines FTS5 full-text search with
Jaccard similarity reranking and trust-weighted scoring.
"""
from __future__ import annotations
import math
from datetime import datetime, timezone
from typing import TYPE_CHECKING
if TYPE_CHECKING:
from .store import MemoryStore
try:
from . import holographic as hrr
except ImportError:
import holographic as hrr # type: ignore[no-redef]
class FactRetriever:
"""Multi-strategy fact retrieval with trust-weighted scoring."""
def __init__(
self,
store: MemoryStore,
temporal_decay_half_life: int = 0, # days, 0 = disabled
fts_weight: float = 0.4,
jaccard_weight: float = 0.3,
hrr_weight: float = 0.3,
hrr_dim: int = 1024,
):
self.store = store
self.half_life = temporal_decay_half_life
self.hrr_dim = hrr_dim
# Auto-redistribute weights if numpy unavailable
if hrr_weight > 0 and not hrr._HAS_NUMPY:
fts_weight = 0.6
jaccard_weight = 0.4
hrr_weight = 0.0
self.fts_weight = fts_weight
self.jaccard_weight = jaccard_weight
self.hrr_weight = hrr_weight
def search(
self,
query: str,
category: str | None = None,
min_trust: float = 0.3,
limit: int = 10,
) -> list[dict]:
"""Hybrid search: FTS5 candidates → Jaccard rerank → trust weighting.
Pipeline:
1. FTS5 search: Get limit*3 candidates from SQLite full-text search
2. Jaccard boost: Token overlap between query and fact content
3. Trust weighting: final_score = relevance * trust_score
4. Temporal decay (optional): decay = 0.5^(age_days / half_life)
Returns list of dicts with fact data + 'score' field, sorted by score desc.
"""
# Stage 1: Get FTS5 candidates (more than limit for reranking headroom)
candidates = self._fts_candidates(query, category, min_trust, limit * 3)
if not candidates:
return []
# Stage 2: Rerank with Jaccard + trust + optional decay
query_tokens = self._tokenize(query)
scored = []
for fact in candidates:
content_tokens = self._tokenize(fact["content"])
tag_tokens = self._tokenize(fact.get("tags", ""))
all_tokens = content_tokens | tag_tokens
jaccard = self._jaccard_similarity(query_tokens, all_tokens)
fts_score = fact.get("fts_rank", 0.0)
# HRR similarity
if self.hrr_weight > 0 and fact.get("hrr_vector"):
fact_vec = hrr.bytes_to_phases(fact["hrr_vector"])
query_vec = hrr.encode_text(query, self.hrr_dim)
hrr_sim = (hrr.similarity(query_vec, fact_vec) + 1.0) / 2.0 # shift to [0,1]
else:
hrr_sim = 0.5 # neutral
# Combine FTS5 + Jaccard + HRR
relevance = (self.fts_weight * fts_score
+ self.jaccard_weight * jaccard
+ self.hrr_weight * hrr_sim)
# Trust weighting
score = relevance * fact["trust_score"]
# Optional temporal decay
if self.half_life > 0:
decay = self._temporal_decay(fact.get("updated_at") or fact.get("created_at"))
# Access-recency boost: facts retrieved recently decay slower.
# A fact accessed within 1 half-life gets up to 1.5x the decay
# factor, tapering to 1.0x (no boost) after 2 half-lives.
last_accessed = fact.get("last_accessed_at")
if last_accessed:
access_boost = self._access_recency_boost(last_accessed)
decay = min(1.0, decay * access_boost)
score *= decay
fact["score"] = score
scored.append(fact)
# Sort by score descending, return top limit
scored.sort(key=lambda x: x["score"], reverse=True)
results = scored[:limit]
# Strip raw HRR bytes — callers expect JSON-serializable dicts
for fact in results:
fact.pop("hrr_vector", None)
return results
def probe(
self,
entity: str,
category: str | None = None,
limit: int = 10,
) -> list[dict]:
"""Compositional entity query using HRR algebra.
Unbinds entity from memory bank to extract associated content.
This is NOT keyword search — it uses algebraic structure to find facts
where the entity plays a structural role.
Falls back to FTS5 search if numpy unavailable.
"""
if not hrr._HAS_NUMPY:
# Fallback to keyword search on entity name
return self.search(entity, category=category, limit=limit)
conn = self.store._conn
# Encode entity as role-bound vector
role_entity = hrr.encode_atom("__hrr_role_entity__", self.hrr_dim)
entity_vec = hrr.encode_atom(entity.lower(), self.hrr_dim)
probe_key = hrr.bind(entity_vec, role_entity)
# Try category-specific bank first, then all facts
if category:
bank_name = f"cat:{category}"
bank_row = conn.execute(
"SELECT vector FROM memory_banks WHERE bank_name = ?",
(bank_name,),
).fetchone()
if bank_row:
bank_vec = hrr.bytes_to_phases(bank_row["vector"])
extracted = hrr.unbind(bank_vec, probe_key)
# Use extracted signal to score individual facts
return self._score_facts_by_vector(
extracted, category=category, limit=limit
)
# Score against individual fact vectors directly
where = "WHERE hrr_vector IS NOT NULL"
params: list = []
if category:
where += " AND category = ?"
params.append(category)
rows = conn.execute(
f"""
SELECT fact_id, content, category, tags, trust_score,
retrieval_count, helpful_count, created_at, updated_at,
hrr_vector
FROM facts
{where}
""",
params,
).fetchall()
if not rows:
# Final fallback: keyword search
return self.search(entity, category=category, limit=limit)
scored = []
for row in rows:
fact = dict(row)
fact_vec = hrr.bytes_to_phases(fact.pop("hrr_vector"))
# Unbind probe key from fact to see if entity is structurally present
residual = hrr.unbind(fact_vec, probe_key)
# Compare residual against content signal
role_content = hrr.encode_atom("__hrr_role_content__", self.hrr_dim)
content_vec = hrr.bind(hrr.encode_text(fact["content"], self.hrr_dim), role_content)
sim = hrr.similarity(residual, content_vec)
fact["score"] = (sim + 1.0) / 2.0 * fact["trust_score"]
scored.append(fact)
scored.sort(key=lambda x: x["score"], reverse=True)
return scored[:limit]
def related(
self,
entity: str,
category: str | None = None,
limit: int = 10,
) -> list[dict]:
"""Discover facts that share structural connections with an entity.
Unlike probe (which finds facts *about* an entity), related finds
facts that are connected through shared context — e.g., other entities
mentioned alongside this one, or content that overlaps structurally.
Falls back to FTS5 search if numpy unavailable.
"""
if not hrr._HAS_NUMPY:
return self.search(entity, category=category, limit=limit)
conn = self.store._conn
# Encode entity as a bare atom (not role-bound — we want ANY structural match)
entity_vec = hrr.encode_atom(entity.lower(), self.hrr_dim)
# Get all facts with vectors
where = "WHERE hrr_vector IS NOT NULL"
params: list = []
if category:
where += " AND category = ?"
params.append(category)
rows = conn.execute(
f"""
SELECT fact_id, content, category, tags, trust_score,
retrieval_count, helpful_count, created_at, updated_at,
hrr_vector
FROM facts
{where}
""",
params,
).fetchall()
if not rows:
return self.search(entity, category=category, limit=limit)
# Score each fact by how much the entity's atom appears in its vector
# This catches both role-bound entity matches AND content word matches
scored = []
for row in rows:
fact = dict(row)
fact_vec = hrr.bytes_to_phases(fact.pop("hrr_vector"))
# Check structural similarity: unbind entity from fact
residual = hrr.unbind(fact_vec, entity_vec)
# A high-similarity residual to ANY known role vector means this entity
# plays a structural role in the fact
role_entity = hrr.encode_atom("__hrr_role_entity__", self.hrr_dim)
role_content = hrr.encode_atom("__hrr_role_content__", self.hrr_dim)
entity_role_sim = hrr.similarity(residual, role_entity)
content_role_sim = hrr.similarity(residual, role_content)
# Take the max — entity could appear in either role
best_sim = max(entity_role_sim, content_role_sim)
fact["score"] = (best_sim + 1.0) / 2.0 * fact["trust_score"]
scored.append(fact)
scored.sort(key=lambda x: x["score"], reverse=True)
return scored[:limit]
def reason(
self,
entities: list[str],
category: str | None = None,
limit: int = 10,
) -> list[dict]:
"""Multi-entity compositional query — vector-space JOIN.
Given multiple entities, algebraically intersects their structural
connections to find facts related to ALL of them simultaneously.
This is compositional reasoning that no embedding DB can do.
Example: reason(["peppi", "backend"]) finds facts where peppi AND
backend both play structural roles — without keyword matching.
Falls back to FTS5 search if numpy unavailable.
"""
if not hrr._HAS_NUMPY or not entities:
# Fallback: search with all entities as keywords
query = " ".join(entities)
return self.search(query, category=category, limit=limit)
conn = self.store._conn
role_entity = hrr.encode_atom("__hrr_role_entity__", self.hrr_dim)
# For each entity, compute what the bank "remembers" about it
# by unbinding entity+role from each fact vector
entity_residuals = []
for entity in entities:
entity_vec = hrr.encode_atom(entity.lower(), self.hrr_dim)
probe_key = hrr.bind(entity_vec, role_entity)
entity_residuals.append(probe_key)
# Get all facts with vectors
where = "WHERE hrr_vector IS NOT NULL"
params: list = []
if category:
where += " AND category = ?"
params.append(category)
rows = conn.execute(
f"""
SELECT fact_id, content, category, tags, trust_score,
retrieval_count, helpful_count, created_at, updated_at,
hrr_vector
FROM facts
{where}
""",
params,
).fetchall()
if not rows:
query = " ".join(entities)
return self.search(query, category=category, limit=limit)
# Score each fact by how much EACH entity is structurally present.
# A fact scores high only if ALL entities have structural presence
# (AND semantics via min, vs OR which would use mean/max).
role_content = hrr.encode_atom("__hrr_role_content__", self.hrr_dim)
scored = []
for row in rows:
fact = dict(row)
fact_vec = hrr.bytes_to_phases(fact.pop("hrr_vector"))
entity_scores = []
for probe_key in entity_residuals:
residual = hrr.unbind(fact_vec, probe_key)
sim = hrr.similarity(residual, role_content)
entity_scores.append(sim)
min_sim = min(entity_scores)
fact["score"] = (min_sim + 1.0) / 2.0 * fact["trust_score"]
scored.append(fact)
scored.sort(key=lambda x: x["score"], reverse=True)
return scored[:limit]
def contradict(
self,
category: str | None = None,
threshold: float = 0.3,
limit: int = 10,
) -> list[dict]:
"""Find potentially contradictory facts via entity overlap + content divergence.
Two facts contradict when they share entities (same subject) but have
low content-vector similarity (different claims). This is automated
memory hygiene — no other memory system does this.
Returns pairs of facts with a contradiction score.
Falls back to empty list if numpy unavailable.
"""
if not hrr._HAS_NUMPY:
return []
conn = self.store._conn
# Get all facts with vectors and their linked entities
where = "WHERE f.hrr_vector IS NOT NULL"
params: list = []
if category:
where += " AND f.category = ?"
params.append(category)
rows = conn.execute(
f"""
SELECT f.fact_id, f.content, f.category, f.tags, f.trust_score,
f.created_at, f.updated_at, f.hrr_vector
FROM facts f
{where}
""",
params,
).fetchall()
if len(rows) < 2:
return []
# Guard against O(n²) explosion on large fact stores.
# At 500 facts, that's ~125K comparisons — acceptable.
# Above that, only check the most recently updated facts.
_MAX_CONTRADICT_FACTS = 500
if len(rows) > _MAX_CONTRADICT_FACTS:
rows = sorted(rows, key=lambda r: r["updated_at"] or r["created_at"], reverse=True)
rows = rows[:_MAX_CONTRADICT_FACTS]
# Build entity sets per fact
fact_entities: dict[int, set[str]] = {}
for row in rows:
fid = row["fact_id"]
entity_rows = conn.execute(
"""
SELECT e.name FROM entities e
JOIN fact_entities fe ON fe.entity_id = e.entity_id
WHERE fe.fact_id = ?
""",
(fid,),
).fetchall()
fact_entities[fid] = {r["name"].lower() for r in entity_rows}
# Compare all pairs: high entity overlap + low content similarity = contradiction
facts = [dict(r) for r in rows]
contradictions = []
for i in range(len(facts)):
for j in range(i + 1, len(facts)):
f1, f2 = facts[i], facts[j]
ents1 = fact_entities.get(f1["fact_id"], set())
ents2 = fact_entities.get(f2["fact_id"], set())
if not ents1 or not ents2:
continue
# Entity overlap (Jaccard)
entity_overlap = len(ents1 & ents2) / len(ents1 | ents2) if (ents1 | ents2) else 0.0
if entity_overlap < 0.3:
continue # Not enough entity overlap to be contradictory
# Content similarity via HRR vectors
v1 = hrr.bytes_to_phases(f1["hrr_vector"])
v2 = hrr.bytes_to_phases(f2["hrr_vector"])
content_sim = hrr.similarity(v1, v2)
# High entity overlap + low content similarity = potential contradiction
# contradiction_score: higher = more contradictory
contradiction_score = entity_overlap * (1.0 - (content_sim + 1.0) / 2.0)
if contradiction_score >= threshold:
# Strip hrr_vector from output (not JSON serializable)
f1_clean = {k: v for k, v in f1.items() if k != "hrr_vector"}
f2_clean = {k: v for k, v in f2.items() if k != "hrr_vector"}
contradictions.append({
"fact_a": f1_clean,
"fact_b": f2_clean,
"entity_overlap": round(entity_overlap, 3),
"content_similarity": round(content_sim, 3),
"contradiction_score": round(contradiction_score, 3),
"shared_entities": sorted(ents1 & ents2),
})
contradictions.sort(key=lambda x: x["contradiction_score"], reverse=True)
return contradictions[:limit]
def _score_facts_by_vector(
self,
target_vec: "np.ndarray",
category: str | None = None,
limit: int = 10,
) -> list[dict]:
"""Score facts by similarity to a target vector."""
conn = self.store._conn
where = "WHERE hrr_vector IS NOT NULL"
params: list = []
if category:
where += " AND category = ?"
params.append(category)
rows = conn.execute(
f"""
SELECT fact_id, content, category, tags, trust_score,
retrieval_count, helpful_count, created_at, updated_at,
hrr_vector
FROM facts
{where}
""",
params,
).fetchall()
scored = []
for row in rows:
fact = dict(row)
fact_vec = hrr.bytes_to_phases(fact.pop("hrr_vector"))
sim = hrr.similarity(target_vec, fact_vec)
fact["score"] = (sim + 1.0) / 2.0 * fact["trust_score"]
scored.append(fact)
scored.sort(key=lambda x: x["score"], reverse=True)
return scored[:limit]
def _fts_candidates(
self,
query: str,
category: str | None,
min_trust: float,
limit: int,
) -> list[dict]:
"""Get raw FTS5 candidates from the store.
Uses the store's database connection directly for FTS5 MATCH
with rank scoring. Normalizes FTS5 rank to [0, 1] range.
"""
conn = self.store._conn
# Build query - FTS5 rank is negative (lower = better match)
# We need to join facts_fts with facts to get all columns
params: list = []
where_clauses = ["facts_fts MATCH ?"]
params.append(query)
if category:
where_clauses.append("f.category = ?")
params.append(category)
where_clauses.append("f.trust_score >= ?")
params.append(min_trust)
where_sql = " AND ".join(where_clauses)
sql = f"""
SELECT f.*, facts_fts.rank as fts_rank_raw
FROM facts_fts
JOIN facts f ON f.fact_id = facts_fts.rowid
WHERE {where_sql}
ORDER BY facts_fts.rank
LIMIT ?
"""
params.append(limit)
try:
rows = conn.execute(sql, params).fetchall()
except Exception:
# FTS5 MATCH can fail on malformed queries — fall back to empty
return []
if not rows:
return []
# Normalize FTS5 rank: rank is negative, lower = better
# Convert to positive score in [0, 1] range
raw_ranks = [abs(row["fts_rank_raw"]) for row in rows]
max_rank = max(raw_ranks) if raw_ranks else 1.0
max_rank = max(max_rank, 1e-6) # avoid div by zero
results = []
for row, raw_rank in zip(rows, raw_ranks):
fact = dict(row)
fact.pop("fts_rank_raw", None)
fact["fts_rank"] = raw_rank / max_rank # normalize to [0, 1]
results.append(fact)
return results
@staticmethod
def _tokenize(text: str) -> set[str]:
"""Simple whitespace tokenization with lowercasing.
Strips common punctuation. No stemming/lemmatization (Phase 1).
"""
if not text:
return set()
# Split on whitespace, lowercase, strip punctuation
tokens = set()
for word in text.lower().split():
cleaned = word.strip(".,;:!?\"'()[]{}#@<>")
if cleaned:
tokens.add(cleaned)
return tokens
@staticmethod
def _jaccard_similarity(set_a: set, set_b: set) -> float:
"""Jaccard similarity coefficient: |A ∩ B| / |A B|."""
if not set_a or not set_b:
return 0.0
intersection = len(set_a & set_b)
union = len(set_a | set_b)
return intersection / union if union > 0 else 0.0
def _temporal_decay(self, timestamp_str: str | None) -> float:
"""Exponential decay: 0.5^(age_days / half_life_days).
Returns 1.0 if decay is disabled or timestamp is missing.
"""
if not self.half_life or not timestamp_str:
return 1.0
try:
if isinstance(timestamp_str, str):
# Parse ISO format timestamp from SQLite
ts = datetime.fromisoformat(timestamp_str.replace("Z", "+00:00"))
else:
ts = timestamp_str
if ts.tzinfo is None:
ts = ts.replace(tzinfo=timezone.utc)
age_days = (datetime.now(timezone.utc) - ts).total_seconds() / 86400
if age_days < 0:
return 1.0
return math.pow(0.5, age_days / self.half_life)
except (ValueError, TypeError):
return 1.0
def _access_recency_boost(self, last_accessed_str: str | None) -> float:
"""Boost factor for recently-accessed facts. Range [1.0, 1.5].
Facts accessed within 1 half-life get up to 1.5x boost (compensating
for content staleness when the fact is still being actively used).
Boost decays linearly to 1.0 (no boost) at 2 half-lives.
Returns 1.0 if half-life is disabled or timestamp is missing.
"""
if not self.half_life or not last_accessed_str:
return 1.0
try:
if isinstance(last_accessed_str, str):
ts = datetime.fromisoformat(last_accessed_str.replace("Z", "+00:00"))
else:
ts = last_accessed_str
if ts.tzinfo is None:
ts = ts.replace(tzinfo=timezone.utc)
age_days = (datetime.now(timezone.utc) - ts).total_seconds() / 86400
if age_days < 0:
return 1.5 # Future timestamp = just accessed
half_lives_since_access = age_days / self.half_life
if half_lives_since_access <= 1.0:
# Within 1 half-life: linearly from 1.5 (just now) to 1.0 (at 1 HL)
return 1.0 + 0.5 * (1.0 - half_lives_since_access)
elif half_lives_since_access <= 2.0:
# Between 1 and 2 half-lives: linearly from 1.0 to 1.0 (no boost)
return 1.0
else:
return 1.0
except (ValueError, TypeError):
return 1.0