docs(google-adk): bring ADK docs to parity with adk-redis 0.0.9 - #3746
docs(google-adk): bring ADK docs to parity with adk-redis 0.0.9#3746nkanu17 wants to merge 10 commits into
Conversation
The redis.io ADK pages were written against adk-redis 0.0.5. Since then 0.0.7 added selectable memory backends, 0.0.8 renamed the session service, and 0.0.9 changed the managed SDK floor and added cache entry IDs. This brings all five pages to 0.0.9. Functional breakage fixed: - The default memory backend is now `redis-agent-memory` (managed), but every sample passed a localhost Agent Memory Server URL with no `backend` override. As written, the quick start silently targeted the managed backend and could not reach the reader's container. All service and tool samples now set `backend` explicitly and the self-hosted path says so. - `create_memory_mcp_toolset` no longer exists in adk-redis. The MCP samples in `redis-agent-memory.md` and `integration-patterns.md` were importing a function that was removed, so they raised ImportError. Both now use ADK's native `McpToolset` + `SseConnectionParams`, matching the fitness_coach_mcp example. - `RedisWorkingMemorySessionService` was renamed to `RedisSessionMemoryService` in 0.0.8. The old name is a deprecated alias that warns and goes away in 0.1.0. All samples use the new name; a note documents the alias and the moved module path. Config tables were wrong in ways a reader would trip over: - `recency_boost` default was `False`, is `True`. - `extraction_strategy` default was `None`, is `discrete`. Added the fourth value, `custom`. - `semantic_weight` / `recency_weight` were 0.7 / 0.3, are 0.8 / 0.2. - `api_base_url` and `default_namespace` were marked Required; both have defaults. - Added the missing `backend`, `api_key`, `store_id`, timeout, threshold, weight, and half-life fields. Also: - New "Choose a memory backend" section with a feature matrix. Recency boost, auto-summarization, extraction strategies, and MCP are self-hosted only. This is the question customers keep asking. - Memory tools: documented all six. `GetMemoryTool` and `MemoryPromptTool` were missing. - Documented invocation-user resolution from `tool_context` and the client-supplied `id` on `CreateMemoryTool` (0.0.9). - Semantic caching: added cache entry IDs, `CacheEntry`, and `delete_by_id()` targeted invalidation (0.0.9). Added the required `server_url` on `LangCacheProviderConfig`. - Examples: nine to ten, added `managed_memory_quickstart`, labeled each memory example with its backend, and noted the runner per example (`python main.py` vs `adk web .`). `redis_search_tools` wires three tools, not four. - Noted the `redis-agent-memory>=0.2.0` floor.
…ated Link the managed service to https://redis.io/agent-memory/ where the two memory backends are introduced, and label the self-hosted `opensource-agent-memory` backend as deprecated. The feature matrix needed a note alongside this. It shows self-hosted as the only backend with recency-boosted search, auto-summarization, extraction strategies, and an MCP endpoint, which without qualification now reads as a recommendation to adopt a deprecated backend. It is framed as current state and a migration-timing consideration instead. Applied wherever the backend choice is presented to the reader: the architecture bullets and prerequisites in `_index.md`, and the backend table and feature matrix in `redis-agent-memory.md`. The per-example backend labels in `examples.md` and `integration-patterns.md` are left as-is; they state which backend an example targets rather than steering the reader toward one.
|
Hi @nkanu17 thank you for this update, last week @n-valchev documented the self-hosted installation of agent memory here, https://redis.io/docs/latest/develop/ai/context-engine/agent-memory/self-managed/ Please update this PR to refer to those new pages instead of the deprecated agent server pages and repo (or in addition to that). Also just fyi, @raphaeldelio is thinking of moving all that self-managed content Nikolay published last week to the /operate section, so you all should coordinate potentially if he does to ensure that link targets are updated. Thank you! |
mich-elle-luna
left a comment
There was a problem hiding this comment.
Looking good, just need to add references to https://redis.io/docs/latest/develop/ai/context-engine/agent-memory/self-managed/ and related content.
… backends Addresses @mich-elle-luna review feedback to reference the new self-managed Agent Memory pages rather than only the deprecated Agent Memory Server repo. Acting on that surfaced a structural problem rather than a missing link. These pages presented a binary: managed means `redis-agent-memory`, self-hosted means `opensource-agent-memory`. That is wrong, and it would send readers to the deprecated backend for the wrong reason. Self-managed Agent Memory serves the same shared Data Plane API as Redis Cloud (`/v1/stores/{storeId}/...`, verified in the self-managed API examples), and `_AgentMemory()` takes the base URL as a positional argument, so it is not Redis Cloud specific. So there are three deployment paths across two backend values: - `redis-agent-memory` + Redis Cloud Data Plane - `redis-agent-memory` + your own self-managed Data Plane - `opensource-agent-memory` + Agent Memory Server (deprecated) You pick a deployment with `api_base_url`, not with `backend`. Readers who want to run Agent Memory themselves should use self-managed with `backend="redis-agent-memory"`, not the deprecated backend. Changes: - New deployment table mapping each path to its `backend`, its `api_base_url`, and its setup guide. - A note warning against reaching for `opensource-agent-memory` merely because a deployment is self-hosted, since the two are different systems rather than two deployments of one system. - Feature matrix columns are now the two backends rather than "Managed" vs "Self-hosted", because the differences follow the backend. Self- managed has the same feature set as Redis Cloud. - Retire "self-hosted" as a synonym for `opensource-agent-memory` across all five pages. It now names Agent Memory Server explicitly, since self-managed Agent Memory is also self-hosted. Links use relref rather than absolute URLs, so if @raphaeldelio moves the self-managed content to /operate the build fails loudly instead of leaving dead links. Verified all 20 relref targets on these pages resolve to existing content files.
|
Thanks @mich-elle-luna, updated in 51c7df9. Acting on this surfaced something bigger than a missing link, so flagging it explicitly for you and @n-valchev. These pages presented a binary: managed means Self-managed Agent Memory serves the same shared Data Plane API as Redis Cloud (
You pick a deployment with
On the One thing worth a second opinion: the feature matrix still shows recency-boosted search, auto-summarization, extraction strategies, and MCP as Agent Memory Server only. I framed that as current state plus a migration-timing consideration, but if there is a roadmap or support timeline for those on the Data Plane API, that wording should say so rather than implying the gap is permanent. |
mich-elle-luna
left a comment
There was a problem hiding this comment.
Hi Nitin, thank you for these updates. Unfortunately, I think this is really confusing with the old deprecated content still weaved into the new info. So, my feedback is to separate or keep all the old info in a section or file that is marked deprecated with a banner, and provide only the info for the supported services that will be maintained going forward in these main files (removing or moving content that really no longer applies. But, I can be persuaded otherwise , if you have perspectives from others that this content must live side-by-side. Thanks!
| 4. The LLM generates a response using session context plus retrieved memories. | ||
| 5. `after_agent_callback` triggers `add_session_to_memory()` for background extraction. | ||
| 6. If the conversation grows long, working memory auto-summarizes older turns. | ||
| 6. On the Agent Memory Server backend, if the conversation grows long, session memory auto-summarizes older turns. |
|
|
||
| Requires prompt engineering to teach the LLM memory management strategy, but gives the agent genuine autonomy over its own memory. | ||
|
|
||
| ### Invocation-scoped users |
| ## MCP tools | ||
|
|
||
| Point ADK's `McpToolset` at the Agent Memory Server's SSE endpoint. Tool discovery happens automatically — no manual tool wiring required. | ||
| Point ADK's native `McpToolset` at the Agent Memory Server's SSE endpoint. Tool discovery happens automatically, so no manual tool wiring is required. |
| Available MCP tools: `search_long_term_memory`, `create_long_term_memories`, `get_long_term_memory`, `edit_long_term_memory`, `delete_long_term_memories`, `memory_prompt`, and `set_working_memory`. | ||
|
|
||
| The most portable approach — swap memory backends without changing agent code. Requires the Agent Memory Server running with MCP support on a separate port. | ||
| This is the most portable approach: swap memory backends without changing agent code. It requires the Agent Memory Server running with MCP support on a separate port. |
| - [Integration patterns]({{< relref "/integrate/google-adk/integration-patterns" >}}): Detailed tradeoff comparison of all three approaches | ||
| - [simple_redis_memory](https://github.com/redis-developer/adk-redis/tree/main/examples/simple_redis_memory): Minimal framework services setup | ||
| - [managed_memory_quickstart](https://github.com/redis-developer/adk-redis/tree/main/examples/managed_memory_quickstart): Redis Agent Memory, no Docker | ||
| - [simple_redis_memory](https://github.com/redis-developer/adk-redis/tree/main/examples/simple_redis_memory): Agent Memory Server with framework services |
Co-authored-by: mich-elle-luna <153109578+mich-elle-luna@users.noreply.github.com>
Co-authored-by: mich-elle-luna <153109578+mich-elle-luna@users.noreply.github.com>
Co-authored-by: mich-elle-luna <153109578+mich-elle-luna@users.noreply.github.com>
Co-authored-by: mich-elle-luna <153109578+mich-elle-luna@users.noreply.github.com>
Sounds good, thanks for the review! |
Per review, the deprecated backend was woven through every page, which made the onboarding path confusing and put unsupported capabilities next to supported ones as if they were peers. All Agent Memory Server content now lives on one page marked with a deprecation warning: setup, the backend-only settings, auto-summarization, extraction strategies, recency boosting, MCP tools, and the three examples written against it. That page also adds a migration path to Redis Agent Memory, which the review asked for. The main pages now cover only supported deployments, Redis Cloud and self-managed, both on backend=redis-agent-memory, and link to the deprecated page for anyone with an existing server: - _index.md: architecture bullet and prerequisites reduced to Redis Agent Memory, Docker setup and backend-switching text removed, deprecation note and a capabilities row pointing at the new page - redis-agent-memory.md: the two-backend comparison is gone. Comparing them feature by feature implied a live choice; the distinction that matters is supported versus deprecated. Config tables drop the backend-only rows - integration-patterns.md: two approaches instead of three, since MCP exists only on the deprecated backend - examples.md: lists the seven examples on supported backends and points to the rest Also applied the reviewer's wording suggestions. One deviation: the suggestion for the installation line dropped the word extra from "the memory extra", which reads as an accidental deletion since it refers to a pip extra, so that word is kept while the deprecated client requirement moves to the deprecated page. Verified every relref target and anchor resolves.
…cated Sharpens the deprecation signal, which was the reviewer's main concern, and cuts the page back so RedisVL docs stop teaching people how to stand up a deprecated server. - The banner now leads with "Deprecated. Do not use for new work." and states that the backend is unmaintained, that its capabilities are not carried forward, and that adk-redis support will be removed. The page says outright that it exists only to help existing deployments migrate - Front matter description and summary both lead with the deprecation, so the status shows in listings and search results rather than only on the page - Removed the Docker setup instructions and the full feature and configuration reference. Those belong to the deprecated server's own repository, which the page links to. What remains is the part specific to adk-redis: which backend value it used, which settings are inert on redis-agent-memory, and a before-and-after migration example - Repointed the anchors in the other pages, since the removed sections took their headings with them Net effect: the deprecated backend is documented in one clearly-labelled place, in about half the space, oriented entirely toward getting off it.
|
Correction to my Aug 3 comment, flagging it rather than quietly editing since @n-valchev was tagged on it. I claimed auto-summarization and extraction strategies were Agent Memory Server only. That was wrong. Redis Agent Memory supports both, on Redis Cloud and self-managed:
I had read an adk-redis config surface as a product capability. @mich-elle-luna your suggested wording on That also removes my "second opinion" question from the earlier comment. I was asking about a roadmap to close a gap that mostly did not exist. On the restructure: Two of your asks are genuinely outside this repo, so flagging rather than silently dropping them:
|
mich-elle-luna
left a comment
There was a problem hiding this comment.
thank you! just one small change for clarity
Co-authored-by: mich-elle-luna <153109578+mich-elle-luna@users.noreply.github.com>
|
The redis.io Google ADK pages were written against adk-redis 0.0.5. Since then 0.0.7 added selectable memory backends, 0.0.8 renamed the session service, and 0.0.9 raised the managed SDK floor and added cache entry IDs. This brings all five pages to 0.0.9, which is the version currently on PyPI.
Following review, the pages now also centre on Redis Agent Memory and quarantine the deprecated
opensource-agent-memorybackend on its own page.Correction to an earlier version of this PR
An earlier revision of this description, and my comment of Aug 3, claimed that auto-summarization and extraction strategies were Agent Memory Server only. That was wrong, and the claim has been removed from the pages.
Redis Agent Memory supports both, on Redis Cloud and self-managed:
discrete,summary,preferences, andcustom, per Agent Memory.The mistake was reading an adk-redis config surface as a product capability.
model_nameandcontext_window_maxare Agent Memory Server passthroughs in adk-redis, but the managed service does this work configured at the service level instead. Of the four capabilities I originally listed, only recency-weighted search (adk-redis'ssemantic_weight/recency_weight) and the MCP endpoint are genuinely Agent Memory Server only.Thanks @mich-elle-luna, whose suggested wording on
_index.mdcaught this.Functional breakage fixed
Two of these were not just stale wording; readers copying the samples would hit failures.
redis-agent-memorythe default backend, but every sample passed a localhost Agent Memory Server URL with nobackendoverride, so the services silently targeted the Data Plane. All samples now setbackendexplicitly.create_memory_mcp_toolsetno longer exists in adk-redis. The MCP samples imported it, so they raisedImportError. The MCP path now uses ADK's nativeMcpToolset+SseConnectionParams, matching thefitness_coach_mcpexample, and lives on the deprecated page since no managed equivalent exists.RedisWorkingMemorySessionServicewas renamed toRedisSessionMemoryServicein 0.0.8. The old name is a deprecated alias that emits aDeprecationWarningand is removed in 0.1.0. All samples use the new name, with a note covering the alias and the moved module path.Config tables corrected
The documented defaults did not match the code:
recency_boostFalseTrueextraction_strategyNonediscrete(plus a fourth value,custom)semantic_weight/recency_weightapi_base_url,default_namespaceAlso added the missing
backend,api_key,store_id, timeout, threshold, weight, and half-life fields.Structure
Per @mich-elle-luna's review, the deprecated backend no longer runs through the main pages:
agent-memory-server.mdis new, carries a deprecation banner, and holds the Docker setup, the backend-only config, the MCP path, migration steps, and the examples still on that backend.api_base_urlrather than bybackend, linking @n-valchev's self-managed pages and the Redis Cloud create-service guide.Other changes
GetMemoryToolandMemoryPromptToolwere missing.tool_contextand the client-suppliedidonCreateMemoryTool(0.0.9).CacheEntry,delete_by_id()targeted invalidation (0.0.9), and the requiredserver_urlonLangCacheProviderConfig.managed_memory_quickstart, and noted the runner per example (python main.pyvsadk web .).redis_search_toolswires three tools, not four.redis-agent-memory>=0.2.0floor.Verification
Every class name, config field, default value, and example was checked against adk-redis
mainat 0.0.9 rather than against the changelog. Product capability claims are checked against the Agent Memory docs in this repo, which is what the correction above came from.Links use
relrefrather than absolute URLs, so if @raphaeldelio moves the self-managed content to/operatethe build fails loudly instead of leaving dead links. All 44relreftargets across these pages resolve to existing content files.I could not get a clean local Hugo build: the same
ByWeighttemplate error reproduces on an unmodifiedmainwith Hugo 0.160.1, so it is a local version mismatch rather than anything in this change. Worth a look at the CI build.Still open, outside this PR
opensource-agent-memory. They live in adk-redis, so redis.io can only label them.fitness_coach_mcpcannot move at all, since MCP has no managed equivalent.RedisWorkingMemorySessionServicewith nobackendfield. It lives outside this repo.🤖 Generated with Claude Code
Note
Low Risk
Documentation-only changes to integration guides; no application code or runtime behavior in this repo.
Overview
Brings the Google ADK integration docs in line with adk-redis 0.0.9 (selectable backends, session service rename, cache entry IDs,
redis-agent-memory>=0.2.0). Redis Agent Memory (backend="redis-agent-memory") is now the default path for new work on Redis Cloud or self-managed deployments, chosen viaapi_base_url,api_key, andstore_idinstead of local Agent Memory Server Docker setup.Adds
agent-memory-server.mdfor the deprecatedopensource-agent-memorybackend with migration steps, backend-only config, and MCP notes; supported pages no longer treat MCP as a first-class pattern. Samples and tables useRedisSessionMemoryService, explicitbackend, and updated config defaults/fields; documents six REST memory tools, invocation-scoped users, and semantic cacheCacheEntry/delete_by_id()plus LangCacheserver_url. Examples page promotesmanaged_memory_quickstartand separates three legacy examples.Reviewed by Cursor Bugbot for commit 6545386. Bugbot is set up for automated code reviews on this repo. Configure here.