MemMachine vs TencentDB Agent Memory

MemMachine 40

Self-hosted rule, with the price lines read for the hosted platform. No x402, MPP or L402 (0 of 40). No hosted prices found on the site or docs (0). Free to self-host with no card (20). An agent can install it from PyPI with no account (20). Total 40.

TencentDB Agent Memory 60

Read with the self-hosted rule. MIT software with nothing to buy from Tencent, so 20 for pricing, 20 for a free start and 20 for use without an account. No x402, MPP or L402 (0). Memory extraction, wiki and code-graph building and the proxied model are paid to whichever LLM provider the owner configures.

MemMachine 79

OpenAPI documents for the server (54 operations) and the hosted platform (12) (25). llms.txt on docs.memmachine.ai with Markdown pages (10). Every operation has a description, and the docs explain memory types and scoping (15 of 20). Typed request models generated from the server, with search options for limits and filters (12 of 15). Examples for each client and the MCP tools, but only 422 is documented on the server and the README's first example doesn't run (8 of 15). /api/v2 paths and tagged releases, no changelog (9 of 15). Total 79.

TencentDB Agent Memory 61

API reading for a self-hosted HTTP service. The Knowledge service has an OpenAPI 3.0.3 file of 28 paths with no security scheme. Memory Core's 109 endpoints and the proxy's routes are described only in Markdown references, and the proxy's main LLM paths are outside them (10 of 25). No website and no llms.txt. The docs are Markdown in the repository and readable raw (5 of 10). Each Core endpoint states what it does, its limits and caveats, such as which deletes skip user checks, and Knowledge's tool list marks its tools read-only. The API references and the changelog are in Chinese; the README and install guide are in English (11 of 20). Zod-checked inputs with bounds (1 to 100 messages, 8,192 characters per message, limit up to 100) and enums for atom types, but the reference records that version comes back as a string from atomic/update and a number from query (10 of 15). Worked request and response examples per endpoint and error tables per module, with the three different error-message formats set out (12 of 15). /v3 paths with v1 and v2 routes listed as deprecated, and a dated public changelog for every release since 0.1.0 (13 of 15).

MemMachine 83

v0.4.0 on PyPI on 8 October 2026, 29 days before the check (30). v0.3.9-post1, v0.4.0, v0.4.1-dev0 and v0.4.1-a1 in 90 days (20). Many open issues and recent reports with few replies (12 of 25). Python and TypeScript clients, both current (12 of 15). uv lock checks, security audit, SBOM and stale-issue workflows (9 of 10). Total 83.

TencentDB Agent Memory 70

Release 2.0.3 on 26 October 2026 by the changelog and its tag, 9 days before this check (30). Six changelog releases in the 90 days from 6 August, 2.0.1-beta.1 (13 August), 2.0.1 (25 August), 2.0.2-beta.1 (7 September), 2.0.2 (9 October), 2.0.3-beta.1 (13 October) and 2.0.3, with image tags v1.0.2 (8 September) and v1.0.3 (22 September) between. The three stable releases 2.0.1, 2.0.2 and 2.0.3 meet the line without the pre-releases, which this run doesn't count (20). GitHub issues and pull requests weren't readable, so reply times weren't judged. The plugin changelog credits fixes to numbered issues and outside pull requests, and the README says issues get a reply within 24 hours, which we couldn't check. A workflow added on 28 October posts an automatic acknowledgement on new external pull requests, which isn't a maintainer reply (10 of 25). Official TypeScript and Python SDKs exist but only as betas, the TypeScript one last published on 14 August 2026, and the Python package name differs between the changelog and its pyproject. No MCP server, so the registry line doesn't apply (6 of 15). npm packages carry registry signatures but no provenance attestations, the deploy pulls :latest images, the one CI workflow skips the tests and doesn't run on the default branch, and 2.0.3 fixed a Knowledge service broken by OpenTelemetry dependency drift (4 of 10).

MemMachine 57

Apache-2.0 (30). The terms say hosted customer data is destroyed on termination, with no period, and use usage data only in aggregate; the privacy policy wasn't read (12 of 30). No deprecation policy or notices found (3 of 20). No telemetry calls found in the server or client source, and the docs don't address it (10 of 20). Total 55.

TencentDB Agent Memory 68

MIT on both branches and in npm metadata, with the codegraph and Hermes Agent code it reuses credited in the README (30 of 30). The README says new chat memory and skills are private by default and sets out what each visibility means. No privacy statement says what each layer keeps or for how long. The proxy records every turn's upstream messages and model replies, the plugin's L0 and L1 retention defaults to never, chat-memory/clear empties L0, L1 and profile content while keeping the asset, and switching storage backends leaves data behind (12 of 30). v1 and v2 routes are marked deprecated with removal 'planned' and no date, and known limitations promise fixes 'in the next version' without dates (6 of 20). The plugin's metrics reporting is off by default and writes to local logs only, and ClickHouse analytics are off by default and documented. Core depends on OpenTelemetry exporters, and we didn't audit whether anything exports by default. The Hermes Docker image's default model endpoint is Tencent Cloud LKE, which the README states (13 of 20).

MemMachine 41

Framework reading for software the owner runs. The server has no authentication and no security scheme in its OpenAPI document (15 of 30). No read-only mode or confirmation for deletes (5 of 20). No guidance on recalled memories as untrusted text (4 of 15). No audit log on the self-hosted server; the hosted console shows usage (5 of 15). SECURITY.md with GitHub private reporting and a two-business-day acknowledgement, with security-audit and SBOM workflows; no security.txt or certification (12 of 20). Total 41.

TencentDB Agent Memory 35

Read as self-hosted software with a network API. Per-user sk-mem- keys, several per user and revocable, and OAuth2 for panel login. Against that, Memory Core's bearer gate is off by default and the .env template says setting it breaks the proxy's auth and session setup, the data plane takes team, agent and user IDs from the caller and needs only x-tdai-service-id, the proxy's rate-limit and session admin routes have no authentication, /v3/instance/destroy rests on the same unset bearer, the deploy README's defaults include an admin key, and the start scripts publish every port on all interfaces (8 of 30). Private-by-default assets with team, restricted and agent visibility, ACLs by user, role and agent, and read-only Knowledge tools. Data-plane deletes and chat-memory/clear make no user-level check and nothing asks for confirmation. Asset reads were scoped to the caller only in 2.0.2, and Agent and Task reads only in an unreleased commit of 28 October (8 of 20). The proxy injects stored memory, skills, imported documents and code into the system prompt, and we found no guidance on prompt injection or memory poisoning. Personal skills are private until shared after review (3 of 15). Participation logs, memory-generation logs and the raw L0 record let the operator trace what was written, and per-call analytics to ClickHouse is available but off by default (9 of 15). CONTRIBUTING.md asks for vulnerability reports by email to agentmemory@tencent.com rather than public issues, with no acknowledgement time, so a disclosure route without a stated time (7). No SECURITY.md on any branch, no security.txt on www.tencent.com, and no bounty or audit named. The npm plugin's postinstall runs a bash script that rewrites OpenClaw's installed files, with output sent to /dev/null. Two access-control fixes in 2026 are under the negative events.