Skip to content

ADR-002: Agent Memory Platform Roadmap (Task 200-225)

Durum

Kabul Edildi → Uygulandi (Tasks 200-225, 2026-04-28)

Roadmap'i 2026-04-24'te onerildi ve onaylandi; v3.0.0 release'i ile tamamlandi. Asagidaki "Sonuc" bolumu hangi taahhutlerin tutuldugunu, hangilerinin v2'ye kaydirildigini belgeliyor.

Baglam

EvolutionDB'nin vizyonu "Powering Long-Term Memory for Agents" — AI agent framework'lerinin kalici hafiza backend'i olarak konumlanmak. Rekabet manzarasi 2026 basi itibari ile:

  • MongoDB — langgraph-store-mongodb paketi ile LangGraph BaseStore/Checkpointer arayuzune direkt backend sunuyor; Atlas Vector Search, Atlas Stream Processing, TTL indexleri, change streams ile reactive akis.
  • Pinecone — vector-only sidecar; filter + metadata ile hybrid search, ancak SQL/JSON yok — uygulamanin iki stack (Postgres + Pinecone) yonetmesi gerekiyor.
  • Zep — managed service; Graphiti temporal knowledge graph (bitemporal edges: valid_from, valid_to, invalid_at); LongMemEval %63.8 (rakiplerin ortalamasi %49).
  • Mem0 — managed ($249/ay Pro), LLM-extracted fact distillation, triple-scoped user/agent/run.
  • Weaviate, Qdrant, LanceDB — vector-first; SQL/ACID yok.

EvoSQL'in mevcut zengin primitif seti bu pivot icin guclu bir temel sunuyor:

  • JSON + arrays (Task 90) — flexible metadata, $.path, JSON_EXTRACT/SET/REMOVE.
  • 7-layer MVCC (Task 37) — snapshot isolation + CLOG + CSN → temporal query icin zaten dogal altyapi, sadece SQL yuzeyi eksik.
  • LISTEN / NOTIFY (Task 91) — commit-time pub/sub; reactive agent workflow'lari icin Mongo Atlas'in change stream polling'ine karsi push-native rakip.
  • Replication + Raft (Task 97) — 3-node HA, synchronous commit, base backup, TLS + auth.
  • RLS (Task 93) — row-level policy → agent multi-tenant namespace isolation icin hazir.
  • Stored procs + triggers — TTL cleanup, auto-summarization trigger surface.
  • TDE (Task 86) — AES-256-CTR page encryption, on-prem / regulated pazar hazir.
  • PG wire protocol — drop-in client uyumu.
  • Histogram + range selectivity + sampled ANALYZE (Task 99-102) — query planner olgun.

Ancak agent-memory backend olabilmek icin kritik 7 eksik var:

  1. Vector tipi + distance fonksiyonlari — VECTOR(N), <=>, <->, <#> yok.
  2. ANN index (HNSW / IVFFLAT) — semantic search icin olmazsa olmaz.
  3. LangGraph-uyumlu native DDL — BaseCheckpointSaver + BaseStore icin SQL-seviyesinde obje tipi.
  4. TTL column + auto-expire daemon — WITH (ttl_column='expires_at') ve arka plan cleanup.
  5. Temporal SQL yuzeyi — MVCC var, FOR SYSTEM_TIME AS OF / WITH SYSTEM VERSIONING syntax'i yok.
  6. Durable streaming subscription — LISTEN/NOTIFY bellek-ucucu; disconnect'te mesaj kaybi.
  7. Native C client library — Python/Go/Rust/Node FFI'nin saglam ortak cekirdegi yok.

Ayrica tek bir framework'le (LangGraph) sinirli kalmamak icin 9 framework'un bellek abstraction'larini karsilayan ortak SQL primitive set'i gerekiyor:

Framework Bellek abstraction'i EvoSQL primitive
LangGraph BaseCheckpointSaver + BaseStore CHECKPOINT + MEMORY STORE
LangChain classic BaseChatMessageHistory, VectorStoreRetrieverMemory, ConversationEntityMemory, ConversationKGMemory MESSAGE LOG + MEMORY STORE + ENTITY STORE + GRAPH STORE
LlamaIndex ChatMemoryBuffer, BaseKVStore, BaseDocumentStore MESSAGE LOG + MEMORY STORE + DOCUMENT STORE
CrewAI ShortTermMemory, LongTermMemory, EntityMemory MEMORY STORE + CHECKPOINT + ENTITY STORE
AutoGen / AG2 Memory protocol + MemoryContent{content,mime,metadata} MEMORY STORE (MIME JSON'da)
Semantic Kernel IMemoryStore, ISemanticTextMemory MEMORY STORE (zaten karsilanir)
Haystack DocumentStore (Mongo-filter dict), ChatMessageStore DOCUMENT STORE + Mongo-filter DSL
Mem0 Memory.add/search/get/update/delete MEMORY STORE + GRAPH STORE (opsiyonel)
Zep Session messages + Graphiti temporal KG MESSAGE LOG + GRAPH STORE (bitemporal)
OpenAI Swarm stateless — caller persists MESSAGE LOG (opsiyonel)

Tek SQL primitive set (7 obje tipi) + thin adapter shim'leri (50-200 satir/adapter) = N framework, tek backend.

Problem ozeti: EvoSQL'in mevcut SQL/MVCC/replication/TDE altyapisi hazir. Agent-memory platformu olmak icin vector + ANN + 7 native DDL objesi + temporal yuzey + durable subscription + C SDK eklemek gerekiyor. Sonrasi hardcoded SQL primitive'leri ve thin Python/Go/Rust adapter'lari ile 9 framework'u karsilamak mumkun.

Karar

Plan dosyasi /Users/wechip/.claude/plans/golden-wobbling-hanrahan.md'de 26 task / 8 sutun olarak bolunmus enterprise push scope'u uygulanacak. Ozet kararlar:

Sutun A — Vector & Semantic Search (Task 200-203)

  • Task 200: VECTOR(N) tipi. Tuple format type_code = 260000 + N, N×4 byte payload, null bitmap tek bit. Parser VECTOR(128) type rule, '[0.1, 0.2, 0.3]'::VECTOR(3) literal, INSERT/SELECT dimension validation.
  • Task 201: cosine_distance, l2_distance, inner_product, l1_distance fonksiyonlari; operatorler <=>, <->, <#> (evovector convention); vector_dim, vector_norm, vector_normalize helpers.
  • Task 202: HNSW ANN index — CREATE INDEX ... USING HNSW (col vector_cosine_ops) WITH (m=16, ef_construction=64). Graph on slotted pages, M-layer linked lists, greedy search + refinement. ORDER BY col <=> $1 LIMIT k planner hook.
  • Task 203: Hybrid search — vector + filter tek pass. Selectivity < %10 → full scan + vector sort; > %10 → HNSW k' = k/selectivity aday + post-filter.

Sutun B — Agent Memory Native DDL/DML (Task 204-206)

  • Task 204: CREATE CHECKPOINT STORE mem_ck DDL; fiziksel tablo __ck_<name>(thread_id, checkpoint_ns, checkpoint_id PK, values JSON, metadata JSON, parent_config JSON, created_at). DML: CHECKPOINT PUT/GET/LIST/PUT WRITES.
  • Task 205: CREATE MEMORY STORE mem WITH (embedding_dim=1536, distance='cosine'). Fiziksel tablo __mem_<name>(namespace TEXT[], key TEXT, value JSON, embedding VECTOR(N), created_at, ttl_at). DML: MEMORY PUT/GET/SEARCH/DELETE/LIST NAMESPACES.
  • Task 206: Namespace hierarchy ((user_id, "memories")) prefix scan + RLS integration — CREATE POLICY ... ON MEMORY STORE mem FOR SEARCH USING (namespace[1] = current_user).

Sutun C — Temporal / Bitemporal Memory (Task 207-209)

  • Task 207: SELECT * FROM memories FOR SYSTEM_TIME AS OF '2026-04-20 14:00'. MVCC timestamp → CSN lookup (CSN ring cache + CLOG binary search). AS OF TIMESTAMP ve AS OF TRANSACTION N variant.
  • Task 208: CREATE TABLE ... WITH SYSTEM VERSIONING — otomatik valid_from, valid_to kolonlari + t_history table; UPDATE/DELETE hook'lari eski satiri history'e tasir.
  • Task 209: SET SYSTEM_TIME_RETENTION = '30 days' — WAL archive pruning + history table pruning daemon.

Sutun D — Reactive Streaming (Task 210-212)

  • Task 210: CREATE SUBSCRIPTION s FOR CHANNEL 'memory_updated' DURABLE — ack-based cursor, disconnect/reconnect'te kaybolan mesajlar resume olur.
  • Task 211: CDC streaming server (GAP-D7 tamamlama) — TCP JSON lines format {"op":"I","table":"memories","pk":"42","new":{...},"lsn":N,"ts":...}. CREATE CDC STREAM DDL + CDC SUBSCRIBE s [FROM LSN n] handshake.
  • Task 212: ON MEMORY PUT mem DO ... — MEMORY STORE seviyesinde trigger; auto-summarization ve audit log sablonlari.

Sutun E — Agent Ops (Task 213-215)

  • Task 213: CREATE TABLE ... WITH (ttl_column='expires_at') + ALTER TABLE ... SET TTL. Background daemon her dakika DELETE FROM t WHERE expires_at < NOW().
  • Task 214: Auto-summarization stored procedure template + token_length(text) built-in + external LLM hook stub.
  • Task 215: CREATE JOB j ON SCHEDULE '*/5 * * * *' DO 'ANALYZE TABLE t' — cron evaluator daemon, SHOW JOBS, DROP JOB.

Sutun F — Framework-Common Primitives (Task 222-225)

  • Task 222: CREATE MESSAGE LOG chat — append-only chat history. LOG APPEND / READ LAST n / TRUNCATE / COUNT; TTL integration.
  • Task 223: CREATE DOCUMENT STORE docs + Mongo-stili JSON filter DSL ({"$and":[{"user_id":"42"},{"score":{"$gt":0.5}}]}) → 9 operator ($eq, $ne, $gt, $gte, $lt, $lte, $in, $nin, $and, $or, $not, $exists). DOCUMENT WRITE / FILTER / SEARCH / DELETE.
  • Task 224: CREATE GRAPH STORE kg — bitemporal knowledge graph (Zep Graphiti parity). Node + edge tablolari, valid_from/valid_to/invalid_at triple-per-edge. GRAPH UPSERT NODE/EDGE, GRAPH NEIGHBORS, GRAPH SEARCH, GRAPH PATH AS OF ts.
  • Task 225: CREATE ENTITY STORE entities — LangChain ConversationEntityMemory + CrewAI EntityMemory sablonu. ENTITY PUT/GET/RANK, mention_count + last_seen tracking.

Sutun G — C Client Library (Task 216-218)

  • Task 216: libevosql-memory.{so,dylib,a} — opaque evo_conn_t, EVO text protocol, checkpoint + memory + vector API, thread-local error handling.
  • Task 217: Streaming API — evo_subscribe(conn, channel, callback, ctx) background thread, CDC client, ack API.
  • Task 218: FFI + multi-framework adapters — LangGraph BaseCheckpointSaver/BaseStore, LangChain BaseChatMessageHistory/VectorStoreRetrieverMemory/ConversationEntityMemory, LlamaIndex BaseKVStore/BaseDocumentStore/ChatMemoryBuffer, CrewAI ShortTerm/LongTerm/Entity, AutoGen Memory, Mem0-compatible REST-benzeri Python API. Go cgo + Rust bindgen stub.

Sutun H — Benchmarks & Positioning (Task 219-221)

  • Task 219: LongMemEval dataset ingest + accuracy/latency karsilastirma (Zep, Mem0, langgraph-store-mongodb, Pinecone). Reactive latency LISTEN/NOTIFY vs Mongo polling. Temporal query AS OF vs Graphiti. Hedef: LongMemEval ≥ Zep (%63.8), p99 checkpoint put < 5 ms.
  • Task 220: 6 framework icin resmi/official compat suite — LangGraph langgraph-checkpoint-tests, LangChain, LlamaIndex, CrewAI, AutoGen, Mem0. CI matrix + async stress (1k concurrent threads).
  • Task 221: README.md rewrite, comparison matrix, architecture diagram, ADR-002 (bu belge), 60-sec quickstart, v3.0.0 major version bump (evolution/db/version.h, deploy/helm/evolutiondb/Chart.yaml), launch blog post.

Sprint Sirasi

Sprint 1: 200-203 (vector foundation) Sprint 2: 204-206 (LangGraph DDL) Sprint 3: 207-209 (temporal) Sprint 4: 210-212 (reactive) Sprint 5: 213-215 (ops) Sprint 6: 222-225 (framework primitives) Sprint 7: 216-218 (C SDK + adapters) Sprint 8: 219-221 (bench + relaunch)

MVP kritik yolu (LangGraph-only pazarlama demosu): 200 → 201 → 202 → 204 → 205 → 216 → 218(LangGraph) → 220 ≈ 3 sprint.

Full enterprise push: 26 task / ~9-10 hafta.

Release: v3.0.0 (major: urun repositioning).

Alternatifler

Secenek 1: Tam enterprise push — 26 task / 8 sutun (bu ADR'in onerdigi karar)

  • Avantajlar: 4 differentiator axis birlikte acilir (SQL+vector+JSON single-process, temporal/bitemporal, reactive push, embedded+TDE on-prem); 9 framework ortak backend; Zep'in tek avantajini (Graphiti) Task 224 ile kopar; LongMemEval %63.8 hedefi; v3.0.0 launch marketing icin yeterli ozellik; C SDK ile Python/Go/Rust/Node FFI tek cekirdek.
  • Dezavantajlar: ~9-10 hafta is, 8 sprint; scope creep riski (mitigation: sprint-level sign-off, MVP yolu 3 sprint icinde demo-ready); Graph store (Task 224) karmasiklik — bitemporal edges + BFS/path implementation bug-prone; HNSW recall tuning 128-dim'de Pinecone ile paritede olmayabilir (mitigation: Task 202 test'inde recall@10 > %90 hard gate).

Secenek 2: MVP-only scope — sadece 200, 201, 202, 204, 205, 216, 218(LG), 220 (~3 sprint)

  • Avantajlar: Hizli demo, LangGraph drop-in calisir, vector search + checkpoint + memory store ile Mongo/Pinecone pazarinin %80'i kapanir; risk az.
  • Dezavantajlar: Tek differentiator axis (single-process SQL+vector); temporal yok → Zep/Graphiti farkini kapayamayiz; reactive/durable subscription yok → Mongo polling'e karsi avantaj sifir; multi-framework coverage yok → LangChain/LlamaIndex/CrewAI/AutoGen/Mem0 kullanicilari adapter yazmak zorunda → tutunma dusuk; Mongo Atlas'in TTL + change streams + temporal olmayisi sasirtici degilse Mongo'yu gecemeyiz. Positioning: "LangGraph-only vector DB" gibi gorunur — agent-memory platformu degil.

Secenek 3: LangGraph-only ama tam feature — 200-215 (sutun F + C SDK + bench yok, ~5-6 sprint)

  • Avantajlar: Temporal + reactive + ops axis'leri dolu; tek framework'e odak; C SDK rotasyonu sprint-8'e ertelenir; riskli tasklar (Graph Store, CDC streaming) sonraya atilir.
  • Dezavantajlar: Multi-framework ekosistem tutunma firsati kacar — LangChain/CrewAI kullanicilari rakiplerde kalir; C SDK yoklugu Python'a bagimli kilar (Go/Rust/Node ekosistemi kendi basina sarmak zorunda); Mem0-compatible REST API yoklugunda Mem0 gocu engellenir; "agent-memory platform" iddiasini kismi tutar.

Secenek 4: Saf vector DB (sadece Sutun A) — Pinecone-lite

  • Avantajlar: Minimal is (~1 sprint); evovector + HNSW + hybrid kombinasyonu mevcut SQL motorlarinda nadir; Postgres vektor eklentilerine alternatif sunar.
  • Dezavantajlar: Agent-memory vizyonu kaybolur — Postgres'in mevcut vektor eklentisi yerini koruyamaz, Qdrant/LanceDB/Weaviate ile rekabet komodite; MVCC/JSON/replication/TDE/RLS avantajlari gorunmez kalir; gelecekte agent platformuna gecis icin yine ayni 26 task'a donmek gerekir.

Secenek 5: External integration — Mongo/Pinecone icin bridge + CDC → "use what exists"

  • Avantajlar: Bakim yuku dista; mevcut Task 91 LISTEN/NOTIFY'yi Pinecone/Mongo'ya akitan adapter yazilir.
  • Dezavantajlar: EvoSQL agent-memory platformu olmaz; MongoDB/Pinecone'a besleyici olur; tum agent-memory pozisyonu rakiplerin pazar lideri konumunu pekistirir; single-process SQL+vector differentiator'i kaybolur; enterprise on-prem kullanicisi icin hic bir avantaj sunmaz (Pinecone managed/Atlas managed).

Sonuclar

Olumlu

  • Tek proses = tek backend: Agent uygulamalari Postgres + Pinecone (+ Mongo session + Redis cache) yerine tek evosql.db ile calisir → operasyonel karmasiklik %70+ duser.
  • ACID + vector + JSON tek TX: Checkpoint write + embedding upsert + metadata update atomik — Mongo + Pinecone dual-write tutarsizligi ortadan kalkar.
  • Temporal primitive native: MVCC zaten var, sadece SQL yuzeyi ekleniyor → AS OF sorgulari sifir ek depolama maliyetinde. Zep Graphiti'nin tek avantaji (LongMemEval %63.8 vs %49) kopyalanabilir.
  • Reaktif push Mongo polling'i yener: LISTEN/NOTIFY + durable subscription + CDC TCP stream ile agent tick-loop latency 10-100x daha dusuk (Mongo 1 s polling interval'a karsi 1-10 ms push).
  • N framework tek backend: 7 SQL primitive + thin adapter shim'leri ile 9 framework (LangGraph, LangChain, LlamaIndex, CrewAI, AutoGen, Semantic Kernel, Haystack, Mem0, Zep) destekli — framework migration vendor lock-in'i kirar.
  • C SDK / FFI: Python'a bagimli degil; Go/Rust/Node/Zig kendi thin wrapper'ini yazabilir; embedded usage senaryosu acik (in-process library modu opsiyonel olarak v3.1'e erteleniyor).
  • On-prem + TDE + HA: AES-256-CTR page encryption + 3-node Raft HA + PBKDF2 auth + TLS → regulated sektorler (finans, saglik) icin Pinecone managed veya Mongo Atlas'a karsi satin alinabilir SKU.
  • pg wire drop-in: Mevcut Postgres tooling'i (DBeaver, pgAdmin, psql, Grafana, DataGrip) agent-memory sorgulari icin direkt calisir — operasyonel ekip icin ogrenme egrisi sifir.
  • Incremental rollout: 8 sprint, her sprint demo-ready deliverable — MVP 3 sprintte launch-ready (LangGraph + vector + checkpoint + memory + C SDK + LangGraph adapter).
  • Mevcut primitive'ler geri donusluyor: bt2_cursor_seek, cat_upsert_stats, Json.c, mvcc_take_snapshot, auto_reclaim_start, notify_registry_init, repl_set_change_callback, tls_wrap_*, conn_t — yeni kod %30-40 mevcut utility uzerine yazilir.

Olumsuz

  • Scope risk: 26 task / 9-10 hafta buyuk bir taahhut; sprint-level sign-off ve MVP yolu (3 sprint) ile risk azaltilir ama urunde "agent-memory" branding'i full sprint'ler bitmeden duyurulmamali — yanlis beklenti yaratir.
  • HNSW recall tuning: 128-dim/1536-dim embedding'lerde Pinecone'un tuned HNSW'si ile paritede olmak deneysel — ilk tests/test_hnsw_index.py'de 10k vector recall@10 > %90 hard gate ile dogrulanmali; altinda kalirsa IVFFLAT fallback.
  • Graph Store (Task 224) karmasiklik: Bitemporal edges + BFS/path/neighbors implementation bug-prone; MVP'den cikarip Sprint 6'ya birakildi; Zep parity icin kritik ama ilk release'de stub + follow-up v3.1 opsiyonu acik.
  • C SDK thread safety: Agent uygulamalari cok-thread'li arar; evo_subscribe callback dispatch + main TX thread + IO thread race kosulu riski (mitigation: adanmis IO thread pattern + tests/test_concurrency.c 1k thread stress).
  • Framework API drift: 6 adapter pinned versiyonlarla baslar; LangChain/LlamaIndex/CrewAI quarterly breaking change yapiyor; CI matrix (Task 220) her framework'u haftalik calistirir ama upgrade debt birikir (mitigation: quarterly upgrade job).
  • Vector bellek patlamasi: 100k × 1536 float = ~600 MB tek tabloda; tuple format VARBINARY'den taskin olabilir → Task 200 streaming encode + overflow page (TOAST-like) gerekir; kisitlama ilk surumde 8 KB tuple limit'i muhtemel.
  • Mongo-filter DSL scope creep: JSON path operatorleri sinirsiz; Task 223 sadece 9 operator ($eq/$ne/$gt/$gte/$lt/$lte/$in/$nin/$and/$or/$not/$exists); gerisi 42709 undefined_operator donulur; kullanici talepleri v2'ye kuyruklanir.
  • Temporal retention disk sisirme: SET SYSTEM_TIME_RETENTION = '30 days' default 7 gun; prune daemon yoksa disk doludurur → Task 209 WAL archive + history table pruning zorunlu; env var EVOSQL_SYSTEM_TIME_RETENTION_DAYS operator knobu.
  • v3.0.0 major bump iletisim maliyeti: 2.x kullanicilari wiki + migration guide'a ihtiyac duyar; agent-memory DDL yeni (breaking degil) ama major bump urun repositioning + SKU soylemi etkiler; launch blog + release notes + pricing page (Task 221) dogru iletisimi yapmali.
  • CDC streaming server bakim yuku: TCP accept loop + per-client WAL decoder + JSON encoding; bug'lar production'da gozlemi gecikmeli olur; opsiyonel flag (--cdc-port 9970) default disabled — opt-in model.
  • C SDK sinirli adapter demo: Task 218 Python adapter'larin hepsi reference implementation; resmi pip paketleri v3.1'de upstream'e (LangGraph contrib) gonderilecek; kullanici ilk gunde kopyala-yapistir kullanir.

Sonuc (Tasks 200-225, 2026-04-28 itibariyle)

Teslim edilenler (✅)

  • Vector + ANN — Task 200 VECTOR(N) tipi, Task 201 distance fonksiyonlari (<=>, <->, <#>), Task 202 HNSW indexi, Task 203 hybrid search.
  • Agent-memory native DDL — Task 204 CHECKPOINT STORE, Task 205 MEMORY STORE, Task 206 namespace hierarchy + RLS.
  • Temporal — Task 207 FOR SYSTEM_TIME AS OF TRANSACTION, Task 208 WITH SYSTEM VERSIONING, Task 209 retention politikasi.
  • Reactive — Task 210 durable subscription cursor, Task 211 CDC streaming, Task 212 MEMORY trigger.
  • Agent ops — Task 213 TTL kolonu, Task 214 auto-summarize procedure template, Task 215 scheduled jobs (v1: takvim + last_run, SQL exec v2'ye Task 170 olarak).
  • Framework primitives — Task 222 MESSAGE LOG, Task 223 DOCUMENT STORE + Mongo filtre DSL'i, Task 224 GRAPH STORE (bitemporal kenarlar), Task 225 ENTITY STORE.
  • C SDK + Python + adapters — Task 216 libevosql-memory core, Task 217 streaming/subscribe, Task 218 Python ctypes binding + 6 framework adapter (LangGraph, LangChain, LlamaIndex, CrewAI, AutoGen, Mem0).
  • Bench + uyumluluk + relaunch — Task 219 latency/reactive/temporal/longmemeval bench harness'i, Task 220 framework-compat CI matrix'i (44 case + stress), Task 221 README rewrite + comparison matrix + quickstart + release 3.0.0.

v2'ye birakilanlar

  • Task 215 SQL exec: scheduled jobs v1 takvim + last_run kayitlari; auto-RECLAIM thread'inden query_execute re-entry'si TLS state'i bozdugu icin SQL bodisi v2'ye (Task 170) birakildi.
  • Task 211 ack kanali: CDC subscriber tarafinda local cursor kaydedildi; sunucu tarafli ack kanali Task 211 v2'de durable subscription queue ile geliyor.
  • Cross-vendor benchmark: Zep / Mem0 / langgraph-store-mongodb / Pinecone karsilastirma satirlari Task 219 v2'de — backend'ler ayri Docker imajlari olarak gelecek.
  • Metadata path filtreleri: Task 223 mongo-filter DSL'i ust seviye kolonlarda calisiyor; meta.subkey JSON path filtreleri v2.

Olculen sonuclar (single-process p99, bench/run_all.py)

op p99
MEMORY PUT ~ 8 ms
MEMORY GET ~ 2 ms
CHECKPOINT PUT ~ 5 ms
MEMORY SEARCH (top-10) ~ 4 ms
NOTIFY push delivery ~ 0.4 ms
polling @ 1 s ~ 990 ms (≈ 2900× yavaslik)

LongMemEval public dataset'i ile cross-vendor karsilastirma v2'ye birakildi; sentetik fixture uzerinde lexical-fallback recall@10 = 1.0 ile harness pipeline'i dogrulandi.

Tek-binary thesis dogrulamasi

Iddia: "MongoDB + Pinecone dual-stack'ini tek binary ile degistir". Sonuc: PG wire (5433) + EVO native (9967) tek evolutiondb/evolutiondb:latest Docker imaji uzerinden:

  • CREATE MEMORY STORE … WITH (embedding_dim=1536) ile vector storage
  • CREATE CHECKPOINT STORE ile LangGraph backend
  • CREATE MESSAGE LOG, CREATE DOCUMENT STORE, CREATE GRAPH STORE, CREATE ENTITY STORE ile diger framework primitive'leri
  • LISTEN/NOTIFY + durable subscription'lar ile push streaming
  • FOR SYSTEM_TIME AS OF TRANSACTION ile time-travel
  • 6 framework adapter'i (Python) ile resmi protokol uyumu (44 compat case + 3 concurrency stress, CI matrix'inde her commit'te yesil)

iddianin altyapisi tamamlandi.

Tarih

  • 2026-04-24 (onerildi)
  • 2026-04-28 (uygulandi, v3.0.0 ile yayina alindi)