hindsight-api 0.3.0__py3-none-any.whl → 0.4.0__py3-none-any.whl

hindsight_api/engine/retain/fact_extraction.py

@@ -114,11 +114,8 @@ class CausalRelation(BaseModel):
     """Causal relationship from this fact to a previous fact (stored format)."""
 
     target_fact_index: int = Field(description="Index of the related fact in the facts array (0-based).")
-    relation_type: Literal["caused_by", "enabled_by", "prevented_by"] = Field(
-        description="How this fact relates to the target: "
-        "'caused_by' = this fact was caused by the target, "
-        "'enabled_by' = this fact was enabled by the target, "
-        "'prevented_by' = this fact was prevented by the target"
+    relation_type: Literal["caused_by"] = Field(
+        description="How this fact relates to the target: 'caused_by' = this fact was caused by the target"
     )
     strength: float = Field(
         description="Strength of relationship (0.0 to 1.0)",
@@ -141,11 +138,8 @@ class FactCausalRelation(BaseModel):
         "MUST be less than this fact's position in the list. "
         "Example: if this is fact #5, target_index can only be 0, 1, 2, 3, or 4."
     )
-    relation_type: Literal["caused_by", "enabled_by", "prevented_by"] = Field(
-        description="How this fact relates to the target fact: "
-        "'caused_by' = this fact was caused by the target fact, "
-        "'enabled_by' = this fact was enabled by the target fact, "
-        "'prevented_by' = this fact was blocked/prevented by the target fact"
+    relation_type: Literal["caused_by"] = Field(
+        description="How this fact relates to the target fact: 'caused_by' = this fact was caused by the target fact"
     )
     strength: float = Field(
         description="Strength of relationship (0.0 to 1.0). 1.0 = strong, 0.5 = moderate",
@@ -438,34 +432,15 @@ def _chunk_conversation(turns: list[dict], max_chars: int) -> list[str]:
 # FACT EXTRACTION PROMPTS
 # =============================================================================
 
-# Concise extraction prompt (default) - selective, high-quality facts
-CONCISE_FACT_EXTRACTION_PROMPT = """Extract SIGNIFICANT facts from text. Be SELECTIVE - only extract facts worth remembering long-term.
+# Base prompt template (shared by concise and custom modes)
+# Uses {extraction_guidelines} placeholder for mode-specific instructions
+_BASE_FACT_EXTRACTION_PROMPT = """Extract SIGNIFICANT facts from text. Be SELECTIVE - only extract facts worth remembering long-term.
 
-LANGUAGE RULE (CRITICAL): Output facts in the EXACT SAME language as the input text. If input is Japanese, output Japanese. If input is Chinese, output Chinese. NEVER translate to English. Preserve original language completely.
+LANGUAGE REQUIREMENT: Detect the language of the input text. All extracted facts, entity names, descriptions, and other output MUST be in the SAME language as the input. Do not translate to another language.
 
 {fact_types_instruction}
 
-══════════════════════════════════════════════════════════════════════════
-SELECTIVITY - CRITICAL (Reduces 90% of unnecessary output)
-══════════════════════════════════════════════════════════════════════════
-
-ONLY extract facts that are:
-✅ Personal info: names, relationships, roles, background
-✅ Preferences: likes, dislikes, habits, interests (e.g., "Alice likes coffee")
-✅ Significant events: milestones, decisions, achievements, changes
-✅ Plans/goals: future intentions, deadlines, commitments
-✅ Expertise: skills, knowledge, certifications, experience
-✅ Important context: projects, problems, constraints
-✅ Sensory/emotional details: feelings, sensations, perceptions that provide context
-✅ Observations: descriptions of people, places, things with specific details
-
-DO NOT extract:
-❌ Generic greetings: "how are you", "hello", pleasantries without substance
-❌ Pure filler: "thanks", "sounds good", "ok", "got it", "sure"
-❌ Process chatter: "let me check", "one moment", "I'll look into it"
-❌ Repeated info: if already stated, don't extract again
-
-CONSOLIDATE related statements into ONE fact when possible.
+{extraction_guidelines}
 
 ══════════════════════════════════════════════════════════════════════════
 FACT FORMAT - BE CONCISE
@@ -513,7 +488,33 @@ ENTITIES
 ══════════════════════════════════════════════════════════════════════════
 
 Include: people names, organizations, places, key objects, abstract concepts (career, friendship, etc.)
-Always include "user" when fact is about the user.
+Always include "user" when fact is about the user.{examples}"""
+
+# Concise mode guidelines
+_CONCISE_GUIDELINES = """══════════════════════════════════════════════════════════════════════════
+SELECTIVITY - CRITICAL (Reduces 90% of unnecessary output)
+══════════════════════════════════════════════════════════════════════════
+
+ONLY extract facts that are:
+✅ Personal info: names, relationships, roles, background
+✅ Preferences: likes, dislikes, habits, interests (e.g., "Alice likes coffee")
+✅ Significant events: milestones, decisions, achievements, changes
+✅ Plans/goals: future intentions, deadlines, commitments
+✅ Expertise: skills, knowledge, certifications, experience
+✅ Important context: projects, problems, constraints
+✅ Sensory/emotional details: feelings, sensations, perceptions that provide context
+✅ Observations: descriptions of people, places, things with specific details
+
+DO NOT extract:
+❌ Generic greetings: "how are you", "hello", pleasantries without substance
+❌ Pure filler: "thanks", "sounds good", "ok", "got it", "sure"
+❌ Process chatter: "let me check", "one moment", "I'll look into it"
+❌ Repeated info: if already stated, don't extract again
+
+CONSOLIDATE related statements into ONE fact when possible."""
+
+# Concise mode examples
+_CONCISE_EXAMPLES = """
 
 ══════════════════════════════════════════════════════════════════════════
 EXAMPLES
@@ -539,6 +540,20 @@ QUALITY OVER QUANTITY
 
 Ask: "Would this be useful to recall in 6 months?" If no, skip it."""
 
+# Assembled concise prompt (backward compatible - exact same output as before)
+CONCISE_FACT_EXTRACTION_PROMPT = _BASE_FACT_EXTRACTION_PROMPT.format(
+    fact_types_instruction="{fact_types_instruction}",
+    extraction_guidelines=_CONCISE_GUIDELINES,
+    examples=_CONCISE_EXAMPLES,
+)
+
+# Custom prompt uses same base but without examples
+CUSTOM_FACT_EXTRACTION_PROMPT = _BASE_FACT_EXTRACTION_PROMPT.format(
+    fact_types_instruction="{fact_types_instruction}",
+    extraction_guidelines="{custom_instructions}",
+    examples="",  # No examples for custom mode
+)
+
 
 # Verbose extraction prompt - detailed, comprehensive facts (legacy mode)
 VERBOSE_FACT_EXTRACTION_PROMPT = """Extract facts from text into structured format with FIVE required dimensions - BE EXTREMELY DETAILED.
@@ -662,7 +677,7 @@ CAUSAL RELATIONSHIPS
 ══════════════════════════════════════════════════════════════════════════
 
 Link facts with causal_relations (max 2 per fact). target_index must be < this fact's index.
-Types: "caused_by", "enabled_by", "prevented_by"
+Type: "caused_by" (this fact was caused by the target fact)
 
 Example: "Lost job → couldn't pay rent → moved apartment"
 - Fact 0: Lost job, causal_relations: null
@@ -686,6 +701,12 @@ async def _extract_facts_from_chunk(
     Note: event_date parameter is kept for backward compatibility but not used in prompt.
     The LLM extracts temporal information from the context string instead.
     """
+    import logging
+
+    from openai import BadRequestError
+
+    logger = logging.getLogger(__name__)
+
     memory_bank_context = f"\n- Your name: {agent_name}" if agent_name and extract_opinions else ""
 
     # Determine which fact types to extract based on the flag
@@ -704,13 +725,27 @@ async def _extract_facts_from_chunk(
     extract_causal_links = config.retain_extract_causal_links
 
     # Select base prompt based on extraction mode
-    if extraction_mode == "verbose":
+    if extraction_mode == "custom":
+        # Custom mode: inject user-provided guidelines
+        if not config.retain_custom_instructions:
+            logger.warning(
+                "extraction_mode='custom' but HINDSIGHT_API_RETAIN_CUSTOM_INSTRUCTIONS not set. "
+                "Falling back to 'concise' mode."
+            )
+            base_prompt = CONCISE_FACT_EXTRACTION_PROMPT
+            prompt = base_prompt.format(fact_types_instruction=fact_types_instruction)
+        else:
+            base_prompt = CUSTOM_FACT_EXTRACTION_PROMPT
+            prompt = base_prompt.format(
+                fact_types_instruction=fact_types_instruction,
+                custom_instructions=config.retain_custom_instructions,
+            )
+    elif extraction_mode == "verbose":
         base_prompt = VERBOSE_FACT_EXTRACTION_PROMPT
+        prompt = base_prompt.format(fact_types_instruction=fact_types_instruction)
     else:
         base_prompt = CONCISE_FACT_EXTRACTION_PROMPT
-
-    # Format the prompt with fact types instruction
-    prompt = base_prompt.format(fact_types_instruction=fact_types_instruction)
+        prompt = base_prompt.format(fact_types_instruction=fact_types_instruction)
 
     # Build the full prompt with or without causal relationships section
     # Select appropriate response schema based on extraction mode and causal links
@@ -723,12 +758,6 @@ async def _extract_facts_from_chunk(
     else:
         response_schema = FactExtractionResponseNoCausal
 
-    import logging
-
-    from openai import BadRequestError
-
-    logger = logging.getLogger(__name__)
-
     # Retry logic for JSON validation errors
     max_retries = 2
     last_error = None
@@ -823,7 +852,8 @@ Text:
 
                 # Critical field: fact_type
                 # LLM uses "assistant" but we convert to "experience" for storage
-                fact_type = llm_fact.get("fact_type")
+                original_fact_type = llm_fact.get("fact_type")
+                fact_type = original_fact_type
 
                 # Convert "assistant" → "experience" for storage
                 if fact_type == "assistant":
@@ -840,7 +870,10 @@ Text:
                     else:
                         # Default to 'world' if we can't determine
                         fact_type = "world"
-                        logger.warning(f"Fact {i}: defaulting to fact_type='world'")
+                        logger.warning(
+                            f"Fact {i}: defaulting to fact_type='world' "
+                            f"(original fact_type={original_fact_type!r}, fact_kind={fact_kind!r})"
+                        )
 
                 # Get fact_kind for temporal handling (but don't store it)
                 fact_kind = llm_fact.get("fact_kind", "conversation")

hindsight_api/engine/retain/fact_storage.py

@@ -41,7 +41,6 @@ async def insert_facts_batch(
     contexts = []
     fact_types = []
     confidence_scores = []
-    access_counts = []
     metadata_jsons = []
     chunk_ids = []
     document_ids = []
@@ -61,7 +60,6 @@ async def insert_facts_batch(
         fact_types.append(fact.fact_type)
         # confidence_score is only for opinion facts
         confidence_scores.append(1.0 if fact.fact_type == "opinion" else None)
-        access_counts.append(0)  # Initial access count
         metadata_jsons.append(json.dumps(fact.metadata))
         chunk_ids.append(fact.chunk_id)
         # Use per-fact document_id if available, otherwise fallback to batch-level document_id
@@ -76,16 +74,16 @@ async def insert_facts_batch(
         WITH input_data AS (
             SELECT * FROM unnest(
                 $2::text[], $3::vector[], $4::timestamptz[], $5::timestamptz[], $6::timestamptz[], $7::timestamptz[],
-                $8::text[], $9::text[], $10::float[], $11::int[], $12::jsonb[], $13::text[], $14::text[], $15::jsonb[]
+                $8::text[], $9::text[], $10::float[], $11::jsonb[], $12::text[], $13::text[], $14::jsonb[]
             ) AS t(text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                   context, fact_type, confidence_score, access_count, metadata, chunk_id, document_id, tags_json)
+                   context, fact_type, confidence_score, metadata, chunk_id, document_id, tags_json)
         )
         INSERT INTO {fq_table("memory_units")} (bank_id, text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                                 context, fact_type, confidence_score, access_count, metadata, chunk_id, document_id, tags)
+                                 context, fact_type, confidence_score, metadata, chunk_id, document_id, tags)
         SELECT
             $1,
             text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-            context, fact_type, confidence_score, access_count, metadata, chunk_id, document_id,
+            context, fact_type, confidence_score, metadata, chunk_id, document_id,
             COALESCE(
                 (SELECT array_agg(elem) FROM jsonb_array_elements_text(tags_json) AS elem),
                 '{{}}'::varchar[]
@@ -103,7 +101,6 @@ async def insert_facts_batch(
         contexts,
         fact_types,
         confidence_scores,
-        access_counts,
         metadata_jsons,
         chunk_ids,
         document_ids,
@@ -126,7 +123,7 @@ async def ensure_bank_exists(conn, bank_id: str) -> None:
     """
     await conn.execute(
         f"""
-        INSERT INTO {fq_table("banks")} (bank_id, disposition, background)
+        INSERT INTO {fq_table("banks")} (bank_id, disposition, mission)
         VALUES ($1, $2::jsonb, $3)
         ON CONFLICT (bank_id) DO UPDATE
         SET updated_at = NOW()

hindsight_api/engine/retain/link_utils.py

@@ -754,17 +754,14 @@ async def create_causal_links_batch(
         causal_relations_per_fact: List of causal relations for each fact.
             Each element is a list of dicts with:
             - target_fact_index: Index into unit_ids for the target fact
-            - relation_type: "causes", "caused_by", "enables", or "prevents"
+            - relation_type: "caused_by"
             - strength: Float in [0.0, 1.0] representing relationship strength
 
     Returns:
         Number of causal links created
 
-    Causal link types:
-    - "causes": This fact directly causes the target fact (forward causation)
-    - "caused_by": This fact was caused by the target fact (backward causation)
-    - "enables": This fact enables/allows the target fact (enablement)
-    - "prevents": This fact prevents/blocks the target fact (prevention)
+    Causal link type:
+    - "caused_by": This fact was caused by the target fact
     """
     if not unit_ids or not causal_relations_per_fact:
         return 0
@@ -787,8 +784,8 @@ async def create_causal_links_batch(
                 relation_type = relation["relation_type"]
                 strength = relation.get("strength", 1.0)
 
-                # Validate relation_type - must match database constraint
-                valid_types = {"causes", "caused_by", "enables", "prevents"}
+                # Validate relation_type - only "caused_by" is supported (DB constraint)
+                valid_types = {"caused_by"}
                 if relation_type not in valid_types:
                     logger.error(
                         f"Invalid relation_type '{relation_type}' (type: {type(relation_type).__name__}) "

hindsight_api/engine/retain/orchestrator.py

@@ -9,7 +9,6 @@ import time
 import uuid
 from datetime import UTC, datetime
 
-from ...config import get_config
 from ..db_utils import acquire_with_retry
 from . import bank_utils
 
@@ -28,9 +27,8 @@ from . import (
     fact_extraction,
     fact_storage,
     link_creation,
-    observation_regeneration,
 )
-from .types import ExtractedFact, ProcessedFact, RetainContent, RetainContentDict
+from .types import EntityLink, ExtractedFact, ProcessedFact, RetainContent, RetainContentDict
 
 logger = logging.getLogger(__name__)
 
@@ -40,7 +38,6 @@ async def retain_batch(
     embeddings_model,
     llm_config,
     entity_resolver,
-    task_backend,
     format_date_fn,
     duplicate_checker_fn,
     bank_id: str,
@@ -59,7 +56,6 @@ async def retain_batch(
         embeddings_model: Embeddings model for generating embeddings
         llm_config: LLM configuration for fact extraction
         entity_resolver: Entity resolver for entity processing
-        task_backend: Task backend for background jobs
         format_date_fn: Function to format datetime to readable string
         duplicate_checker_fn: Function to check for duplicate facts
         bank_id: Bank identifier
@@ -408,27 +404,9 @@ async def retain_batch(
             causal_link_count = await link_creation.create_causal_links_batch(conn, unit_ids, non_duplicate_facts)
             log_buffer.append(f"[10] Causal links: {causal_link_count} links in {time.time() - step_start:.3f}s")
 
-            # Regenerate observations - sync (in transaction) or async (background task)
-            config = get_config()
-            if config.retain_observations_async:
-                # Queue for async processing after transaction commits
-                entity_ids_for_async = list(set(link.entity_id for link in entity_links)) if entity_links else []
-                log_buffer.append(
-                    f"[11] Observations: queued {len(entity_ids_for_async)} entities for async processing"
-                )
-            else:
-                # Run synchronously inside transaction for atomicity
-                await observation_regeneration.regenerate_observations_batch(
-                    conn, embeddings_model, llm_config, bank_id, entity_links, log_buffer
-                )
-                entity_ids_for_async = []
-
             # Map results back to original content items
             result_unit_ids = _map_results_to_contents(contents, extracted_facts, is_duplicate_flags, unit_ids)
 
-        # Trigger background tasks AFTER transaction commits
-        await _trigger_background_tasks(task_backend, bank_id, unit_ids, non_duplicate_facts, entity_ids_for_async)
-
         # Log final summary
         total_time = time.time() - start_time
         log_buffer.append(f"{'=' * 60}")
@@ -470,35 +448,3 @@ def _map_results_to_contents(
         result_unit_ids.append(content_unit_ids)
 
     return result_unit_ids
-
-
-async def _trigger_background_tasks(
-    task_backend,
-    bank_id: str,
-    unit_ids: list[str],
-    facts: list[ProcessedFact],
-    entity_ids_for_observations: list[str] | None = None,
-) -> None:
-    """Trigger background tasks after transaction commits."""
-    # Trigger opinion reinforcement if there are entities
-    fact_entities = [[e.name for e in fact.entities] for fact in facts]
-    if any(fact_entities):
-        await task_backend.submit_task(
-            {
-                "type": "reinforce_opinion",
-                "bank_id": bank_id,
-                "created_unit_ids": unit_ids,
-                "unit_texts": [fact.fact_text for fact in facts],
-                "unit_entities": fact_entities,
-            }
-        )
-
-    # Trigger observation regeneration if async mode is enabled
-    if entity_ids_for_observations:
-        await task_backend.submit_task(
-            {
-                "type": "regenerate_observations",
-                "bank_id": bank_id,
-                "entity_ids": entity_ids_for_observations,
-            }
-        )

hindsight_api/engine/retain/types.py

@@ -86,10 +86,10 @@ class CausalRelation:
     """
     Causal relationship between facts.
 
-    Represents how one fact causes, enables, or prevents another.
+    Represents how one fact was caused by another.
     """
 
-    relation_type: str  # "causes", "enables", "prevents", "caused_by"
+    relation_type: str  # "caused_by"
     target_fact_index: int  # Index of the target fact in the batch
     strength: float = 1.0  # Strength of the causal relationship
 

hindsight_api/engine/search/graph_retrieval.py

@@ -162,7 +162,7 @@ class BFSGraphRetriever(GraphRetriever):
         entry_points = await conn.fetch(
             f"""
             SELECT id, text, context, event_date, occurred_start, occurred_end,
-                   mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, tags,
+                   mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
                    1 - (embedding <=> $1::vector) AS similarity
             FROM {fq_table("memory_units")}
             WHERE bank_id = $2
@@ -216,7 +216,7 @@ class BFSGraphRetriever(GraphRetriever):
                 neighbors = await conn.fetch(
                     f"""
                     SELECT mu.id, mu.text, mu.context, mu.occurred_start, mu.occurred_end,
-                           mu.mentioned_at, mu.access_count, mu.embedding, mu.fact_type,
+                           mu.mentioned_at, mu.embedding, mu.fact_type,
                            mu.document_id, mu.chunk_id, mu.tags,
                            ml.weight, ml.link_type, ml.from_unit_id
                     FROM {fq_table("memory_links")} ml

hindsight_api/engine/search/link_expansion_retrieval.py

@@ -45,7 +45,7 @@ async def _find_semantic_seeds(
     rows = await conn.fetch(
         f"""
         SELECT id, text, context, event_date, occurred_start, occurred_end,
-               mentioned_at, access_count, embedding, fact_type, document_id, chunk_id, tags,
+               mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
                1 - (embedding <=> $1::vector) AS similarity
         FROM {fq_table("memory_units")}
         WHERE bank_id = $2
@@ -155,7 +155,6 @@ class LinkExpansionRetriever(GraphRetriever):
                 all_seeds.extend(temporal_seeds)
 
             if not all_seeds:
-                logger.debug("[LinkExpansion] No seeds found, returning empty results")
                 return [], timings
 
             seed_ids = list({s.id for s in all_seeds})
@@ -164,36 +163,108 @@ class LinkExpansionRetriever(GraphRetriever):
             # Run entity and causal expansion sequentially on same connection
             query_start = time.time()
 
-            entity_rows = await conn.fetch(
-                f"""
-                SELECT
-                    mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                    mu.occurred_end, mu.mentioned_at, mu.access_count, mu.embedding,
-                    mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
-                    COUNT(*)::float AS score
-                FROM {fq_table("unit_entities")} seed_ue
-                JOIN {fq_table("entities")} e ON seed_ue.entity_id = e.id
-                JOIN {fq_table("unit_entities")} other_ue ON seed_ue.entity_id = other_ue.entity_id
-                JOIN {fq_table("memory_units")} mu ON other_ue.unit_id = mu.id
-                WHERE seed_ue.unit_id = ANY($1::uuid[])
-                  AND e.mention_count < $2
-                  AND mu.id != ALL($1::uuid[])
-                  AND mu.fact_type = $3
-                GROUP BY mu.id
-                ORDER BY score DESC
-                LIMIT $4
-                """,
-                seed_ids,
-                self.max_entity_frequency,
-                fact_type,
-                budget,
-            )
+            # For observations, traverse through source_memory_ids to find entity connections.
+            # Observations don't have direct unit_entities - they inherit entities via their
+            # source world/experience facts.
+            #
+            # Path: observation → source_memory_ids → world fact → entities →
+            #       ALL world facts with those entities → their observations (excluding seeds)
+            if fact_type == "observation":
+                # Debug: Check what source_memory_ids exist on seed observations
+                debug_sources = await conn.fetch(
+                    f"""
+                    SELECT id, source_memory_ids
+                    FROM {fq_table("memory_units")}
+                    WHERE id = ANY($1::uuid[])
+                    """,
+                    seed_ids,
+                )
+                source_ids_found = []
+                for row in debug_sources:
+                    if row["source_memory_ids"]:
+                        source_ids_found.extend(row["source_memory_ids"])
+                logger.debug(
+                    f"[LinkExpansion] observation graph: {len(seed_ids)} seeds, "
+                    f"{len(source_ids_found)} source_memory_ids found"
+                )
+
+                entity_rows = await conn.fetch(
+                    f"""
+                    WITH seed_sources AS (
+                        -- Get source memory IDs from seed observations
+                        SELECT DISTINCT unnest(source_memory_ids) AS source_id
+                        FROM {fq_table("memory_units")}
+                        WHERE id = ANY($1::uuid[])
+                          AND source_memory_ids IS NOT NULL
+                    ),
+                    source_entities AS (
+                        -- Get entities from those source memories (filtered by frequency)
+                        SELECT DISTINCT ue.entity_id
+                        FROM seed_sources ss
+                        JOIN {fq_table("unit_entities")} ue ON ss.source_id = ue.unit_id
+                        JOIN {fq_table("entities")} e ON ue.entity_id = e.id
+                        WHERE e.mention_count < $2
+                    ),
+                    all_connected_sources AS (
+                        -- Find ALL world facts sharing those entities (don't exclude seed sources)
+                        -- The exclusion happens at the observation level, not the source level
+                        SELECT DISTINCT other_ue.unit_id AS source_id
+                        FROM source_entities se
+                        JOIN {fq_table("unit_entities")} other_ue ON se.entity_id = other_ue.entity_id
+                    )
+                    -- Find observations derived from connected source memories
+                    -- Only exclude the actual seed observations
+                    SELECT
+                        mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
+                        mu.occurred_end, mu.mentioned_at, mu.embedding,
+                        mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
+                        COUNT(DISTINCT cs.source_id)::float AS score
+                    FROM all_connected_sources cs
+                    JOIN {fq_table("memory_units")} mu
+                        ON mu.source_memory_ids @> ARRAY[cs.source_id]
+                    WHERE mu.fact_type = 'observation'
+                      AND mu.id != ALL($1::uuid[])
+                    GROUP BY mu.id
+                    ORDER BY score DESC
+                    LIMIT $3
+                    """,
+                    seed_ids,
+                    self.max_entity_frequency,
+                    budget,
+                )
+                logger.debug(f"[LinkExpansion] observation graph: found {len(entity_rows)} connected observations")
+            else:
+                # For world/experience facts, use direct entity lookup
+                entity_rows = await conn.fetch(
+                    f"""
+                    SELECT
+                        mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
+                        mu.occurred_end, mu.mentioned_at, mu.embedding,
+                        mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
+                        COUNT(*)::float AS score
+                    FROM {fq_table("unit_entities")} seed_ue
+                    JOIN {fq_table("entities")} e ON seed_ue.entity_id = e.id
+                    JOIN {fq_table("unit_entities")} other_ue ON seed_ue.entity_id = other_ue.entity_id
+                    JOIN {fq_table("memory_units")} mu ON other_ue.unit_id = mu.id
+                    WHERE seed_ue.unit_id = ANY($1::uuid[])
+                      AND e.mention_count < $2
+                      AND mu.id != ALL($1::uuid[])
+                      AND mu.fact_type = $3
+                    GROUP BY mu.id
+                    ORDER BY score DESC
+                    LIMIT $4
+                    """,
+                    seed_ids,
+                    self.max_entity_frequency,
+                    fact_type,
+                    budget,
+                )
 
             causal_rows = await conn.fetch(
                 f"""
                 SELECT DISTINCT ON (mu.id)
                     mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
-                    mu.occurred_end, mu.mentioned_at, mu.access_count, mu.embedding,
+                    mu.occurred_end, mu.mentioned_at, mu.embedding,
                     mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
                     ml.weight + 1.0 AS score
                 FROM {fq_table("memory_links")} ml
@@ -211,11 +282,69 @@ class LinkExpansionRetriever(GraphRetriever):
                 budget,
             )
 
+            # Fallback: semantic/temporal/entity links from memory_links table
+            # These are secondary to entity links (via unit_entities) and causal links
+            # Weight is halved (0.5x) to prioritize primary link types
+            # Check both directions: seeds -> others AND others -> seeds
+            fallback_rows = await conn.fetch(
+                f"""
+                WITH outgoing AS (
+                    -- Links FROM seeds TO other facts
+                    SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
+                           mu.occurred_end, mu.mentioned_at, mu.embedding,
+                           mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
+                           ml.weight
+                    FROM {fq_table("memory_links")} ml
+                    JOIN {fq_table("memory_units")} mu ON ml.to_unit_id = mu.id
+                    WHERE ml.from_unit_id = ANY($1::uuid[])
+                      AND ml.link_type IN ('semantic', 'temporal', 'entity')
+                      AND ml.weight >= $2
+                      AND mu.fact_type = $3
+                      AND mu.id != ALL($1::uuid[])
+                ),
+                incoming AS (
+                    -- Links FROM other facts TO seeds (reverse direction)
+                    SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start,
+                           mu.occurred_end, mu.mentioned_at, mu.embedding,
+                           mu.fact_type, mu.document_id, mu.chunk_id, mu.tags,
+                           ml.weight
+                    FROM {fq_table("memory_links")} ml
+                    JOIN {fq_table("memory_units")} mu ON ml.from_unit_id = mu.id
+                    WHERE ml.to_unit_id = ANY($1::uuid[])
+                      AND ml.link_type IN ('semantic', 'temporal', 'entity')
+                      AND ml.weight >= $2
+                      AND mu.fact_type = $3
+                      AND mu.id != ALL($1::uuid[])
+                ),
+                combined AS (
+                    SELECT * FROM outgoing
+                    UNION ALL
+                    SELECT * FROM incoming
+                )
+                SELECT DISTINCT ON (id)
+                    id, text, context, event_date, occurred_start,
+                    occurred_end, mentioned_at, embedding,
+                    fact_type, document_id, chunk_id, tags,
+                    (MAX(weight) * 0.5) AS score
+                FROM combined
+                GROUP BY id, text, context, event_date, occurred_start,
+                         occurred_end, mentioned_at, embedding,
+                         fact_type, document_id, chunk_id, tags
+                ORDER BY id, score DESC
+                LIMIT $4
+                """,
+                seed_ids,
+                self.causal_weight_threshold,
+                fact_type,
+                budget,
+            )
+
             timings.edge_load_time = time.time() - query_start
-            timings.db_queries = 2
-            timings.edge_count = len(entity_rows) + len(causal_rows)
+            timings.db_queries = 3
+            timings.edge_count = len(entity_rows) + len(causal_rows) + len(fallback_rows)
 
         # Merge results, taking max score per fact
+        # Priority: entity links (unit_entities) > causal links > fallback links
         score_map: dict[str, float] = {}
         row_map: dict[str, dict] = {}
 
@@ -230,6 +359,12 @@ class LinkExpansionRetriever(GraphRetriever):
             if fact_id not in row_map:
                 row_map[fact_id] = dict(row)
 
+        for row in fallback_rows:
+            fact_id = str(row["id"])
+            score_map[fact_id] = max(score_map.get(fact_id, 0), row["score"])
+            if fact_id not in row_map:
+                row_map[fact_id] = dict(row)
+
         # Sort by score and limit
         sorted_ids = sorted(score_map.keys(), key=lambda x: score_map[x], reverse=True)[:budget]
         rows = [row_map[fact_id] for fact_id in sorted_ids]