Dakera vs pgvector: you already run Postgres, so here is what you would still write

pgvector turns Postgres into a vector store, and for vectors that belong next to relational data it is hard to beat. Agent memory asks for more than nearest-neighbour search. This page puts the SQL and application code for hybrid search, reranking and decay beside the Dakera call that does the same.

Reviewed October 2026

When pgvector is exactly right

pgvector is an open-source Postgres extension under the PostgreSQL licence. It stores vectors with the rest of your data and supports exact and approximate search with HNSW and IVFFlat indexes, plus half-precision, binary and sparse vector types and six distance operators. Because it lives in Postgres you get ACID transactions, joins and point-in-time recovery, and it comes preinstalled with many hosted Postgres providers. If your embeddings describe rows you already keep in Postgres, adding a column is the simplest design.

What you would still write

For agent memory, the vector index is the first of several pieces. pgvector's own documentation points to Postgres full-text search and Reciprocal Rank Fusion for hybrid retrieval; the rest is your code.

Write path: schema, insert, decay

The pgvector schema is valid, ordinary Postgres. Decay becomes a scheduled DELETE that you own and tune.

CREATE EXTENSION vector;

CREATE TABLE memories (
  id bigserial PRIMARY KEY,
  agent_id text NOT NULL,
  content text NOT NULL,
  importance real DEFAULT 0.5,
  created_at timestamptz DEFAULT now(),
  embedding vector(1024),
  tsv tsvector GENERATED ALWAYS AS
    (to_tsvector('english', content))
    STORED
);
CREATE INDEX ON memories USING hnsw
  (embedding vector_cosine_ops);
CREATE INDEX ON memories USING gin (tsv);

INSERT INTO memories (agent_id, content,
  importance, embedding)
VALUES ($1, $2, $3, $4);

-- decay: a job you schedule yourself
DELETE FROM memories
WHERE importance * exp(-extract(epoch
  FROM now() - created_at) / 604800.0)
  < 0.01;

from dakera import DakeraClient

client = DakeraClient(
    "http://localhost:3000", api_key=KEY)

client.store_memory(
    agent_id="support",
    content="Maria prefers email",
    importance=0.8,
)

# decay is server-side configuration
client.decay_update_config(
    strategy="exponential",
    half_life_hours=168,
    min_importance=0.01,
)

Read path: hybrid search, rerank, decay

Left: a reciprocal-rank-fusion query over vector and full-text results, weighted by importance and age, then a cross-encoder run in your application. Right: the call that does it.

WITH v AS (
  SELECT id, row_number() OVER (
    ORDER BY embedding <=> $1) AS r
  FROM memories WHERE agent_id = $2
  ORDER BY embedding <=> $1 LIMIT 50),
k AS (
  SELECT id, row_number() OVER (
    ORDER BY ts_rank_cd(tsv, q) DESC) AS r
  FROM memories,
    websearch_to_tsquery('english', $3) q
  WHERE agent_id = $2 AND tsv @@ q
  ORDER BY ts_rank_cd(tsv, q) DESC
  LIMIT 50)
SELECT m.id, m.content,
  (COALESCE(1.0 / (60 + v.r), 0)
   + COALESCE(1.0 / (60 + k.r), 0))
  * m.importance
  * exp(-extract(epoch FROM now()
      - m.created_at) / 604800.0) AS score
FROM memories m
LEFT JOIN v USING (id)
LEFT JOIN k USING (id)
WHERE v.id IS NOT NULL OR k.id IS NOT NULL
ORDER BY score DESC LIMIT 20;

# then, in your application:
rows = conn.execute(SQL, params).fetchall()
pairs = [(query, r[1]) for r in rows]
scores = reranker.predict(pairs)
top = sorted(zip(scores, rows),
             key=lambda x: -x[0])[:10]

response = client.recall(
    agent_id="support",
    query="How do we contact Maria?",
    top_k=10,
)
for m in response.memories:
    print(m.content, m.score)

# BM25 + HNSW, RRF fusion, cross-encoder
# rerank and decay-aware scoring all ran
# in that one call, on-device.

What v0.12.0 adds for Postgres teams

• Multilingual: the SQL above uses one text search configuration, 'english', and you pick another per language. Dakera embeds with bge-m3 across 100+ languages in one space and takes a per-request lang. See Multilingual.
• Multimodal: file attachments, on-device speech to text and visual recall over document pages. See Multimodal.
• Debuggable ranking: rerank_report in every response, plus a Recall lab in the dashboard that shows why each memory ranked where it did.
• Operations: encrypted, compressed backups, /health/ready, and dakera downgrade to roll back to v0.11.108.

When to choose pgvector, and when Dakera

Prefer managed? Dakera Cloud is coming. Join the waitlist.

Frequently asked questions

Related comparisons: vs Qdrant · vs Redis · vs ChromaDB · Mem0 self-hosting